ࡱ> (&%$#! "|0DTimes00BBHpPHp1<| ` .  @n?" dd@  @@`` |d    HH   `b$khLvH>ߨ{@b$<;0_po;YF{@b$AX(x 1!L@b$> $8!^.0*@b$KeQYʴ ъ97@$@b$~ x K݂O@b$Cč0:k , ,4@S @8 -6uʚ;J.ʚ;g4ededQ``ppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 40___PPT10HA-ߌa7ƇR<;0_po;YPNG  IHDR@$JgAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME 9"tN36 IDATxTם/wdS@Y iql֫ncȱ!IT׼v Zw4Y$xic4GJM,kTbFz,u%G:7$Aff$ӆ0fޙg y/R'C PHTR*)x <@ @ PHTR*)x <@n{5DG+Yp8khLvH>ߨPNG  IHDR@֛AgAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME /9T IDATxy@M/*E -,fVٲo=L2e #ˌ!Lv[+"vݹss]ν~v9y9|] ( B'BP(P!P( EBNP(CBP(= 9BP(z r BP* B1T) Bca+w( Ba9BP(z"AP(Fi[rQYf>/\ǜP( "EQK*w( BQEq[e B5JvP( EBNP(CBP(= 9BP(z r BP* B;v숍; :BP(Z|7>r'*WQr$7eBQ'III%JPz]% Eׯ_WVԔ^V(򢸲۝;wfnܸ!grCHsԩm&&&޾}ŅWP,P(jĉ;vomPt yVNBEQ#/RL) ZXX~ԔP(y𡏏ϋ/*Uddd)@Z陚={~T) lٲ.];GK̛g͕b-13qob&+T4/w3OD&BYH9L[Ջ. y{)؉d_ďN=1wq作 dBX3\:.S LfS(pƍ~~"55̌ߌ8t⍼+xL:u@p8zCc9&.] qB_'݅;\;{a[9z)>>^mAhD6-Z*Ne$zJ,٤I؀sUhA!G rebTF/i3 vZzam7P y3K mW>N3f7u ZI!A!fLmSTs,="ٮ})J 90!1c߹P/ݻWKy~R!Coܸj_Da~ŧ߾"(#H};i\Y pvUӹP+VfM3B*^nNzz*55ښtqԺL7(2|Ve GLSu@`QDo Eh:Ex&(r`^T!/ca"HX<'6.8p2n;@ <ԛ6y9$]R4BN}s@' ~7=,[lY3OЕ+Dޙ۽3a-0?ѣZΕ~a`ZZfitsRJтSt"PQS( StBNՂ5ih۶蛣G ;YYYo޼16VЛDɫWԩRZ۷vvvtiBu0n8@JlllƍSt-Z!..I&X!( yv*]0`S!WGN%tr@U 9|2--M>~ >F)S!P۷o+WJ*|gB Dʊ 6|YZZs/kH+BB BStV*/޲eBknIQFb 5T)Ñn" }znnKZhċE)O>YXX)SP(A)-_˗/@˟={ьy?ùJ(FCBU\xɉU)Y&*))Is6Q!P EVJɖȏ.њKB'h9EaAN(o޼INNPB222lmm333 vGݻw+W~_~]vm`4=%**jΝv:#P W|||StӧO7i҄KK*U\vMM7Q!PڮN} k:r T)* '%%}vښDQl }7o9Pü-[r6KnРٳg՛+~B-9/t֍\P(P@Z ~߾}3Y~FBCR EI#rz3r!&zmŊ5z*s@K(J"o߾UWE;K1R!P2+EQ]x{{#)dw3 9БnG]fi]իWhBT)]Ç׮]ӂ9P8۷X#rA!3gθ*UJBN:O411;#\_^reKKKՓSb'OhG 9vur111J/Vs[w^(БnG=}] *ۘZNqRSS֭!Eg铅ʪZvvvF'F[ٹs篿zՓRiQ\ٍ7E8qDVSt .KŁe|||ՕfCNC;)ٳվW=\4ILL-9G,X\P(r| I[S޻w@D"enfzzjzP#G(QB-SztSN!;#FԪU|ErC yեfM&3qob&+T4/w/JPo߾ 4(55U|xf!`͚5lʢҥK?~jժ7nݨQ#cccm^TƲ@p:vH (mb {rJDϞ=|>={AEU~{{{Oǻ xZQoܺq`Roti琐zrP≉A$P!/|rŊ/^?OKIIy5/x٫PwܰmkB`M.(BCܶm5jH|Oyɒχqj /#GYn]/ޢE }/?իWEwНԼSYic{ EOKK۽{O@o߾maaQR%3BN\\!^ܹseP[ hBr٘V^@yS(*ieeuU-\ G]L_7olh!';(Ǐj׮]ZdSą M)mq%WW׀iӦir&LVkRׯy-DCP!wo۷ؐyԩSÇO>ftmn]*Knv_dAS(: "ŋwyݺu+Wv횦wxڵk-ZQN[׿~z* y̙pmQ/ٺu'Oj-cCbbijմi WaCNN_̜ǮϚaÆi}!==ׯ_\@ٳg5w˓\ɓ'k.ΐp$44+:::>>^u}Сɘxuf<3qӶ̌(dҵeD?^Ily j4h%J _PXsBt#G(>QFǭxuӧk׮ӧO^lmm=k,dL8qvȺQOs4quh)ׯ_͛gee(ŋ4wiWWWGGG&G™8pr9`ᩐ;w$ +wlnns&_ΕNgSSӌ M$T4:ͭ]GD50lT[U:Ɯz 8>C%~HNN~&Zcbb6WHۀ-ӧ˗/Msۗt6)שSGy.]TvmZ4ӴnZMSz&[0fhϪ 0`ѳgϖ9֬D>>>PО={D 7DCsrr23SE7Q5f̘qgA?n8Mృ\nmfĶQb{9/_ mjIZFh>,,~غu={Nj$VY&==}ǎ wss{ujj>{D'/7o0d*yuChi@XbJϭ[\yENPw6n8vXsss)\p!/BIfffxxw}c46 u2nixDN)@njp<::xG/_&6Ν;3L B"<ǵk^HRKou@q|‚O>ueԩUV=tPrr2Dʕ+a}U )_|ZZylDԣ`I)EɓBٱcG&qctyʕܾ}[ż,"BEjװԣE7jHbŊаsI^^^ǎ߽{i&y93͛E8'h~q9s\2i$ĚzE޽{NNI4'N͹~zR4hp9???~yoI"b#FTXqڵ .Zd SY8xРA̯affgn8p 4!A"K.U=M`fNK:\7A%o߾maaqlڴiLL&*U ϟ:u.TN ׆͘0bČ'｟^Ǐ,XPN2eܿnݺnnn.]6mF]MʕC$C>CCCq-+++V۷tzMNz.tـ"(XjKnFӮ];(ŊVE-[o߾sԐ·r~%E×Vn%~yf{~qJJʐ!C*Wj0Ϙ1~`ս{VXɓ'/^DchhPm@; ҥiDk7cǎ߿wqqYx1 Yj*ɞ>}C]N:XBJʥو<Ҧ3 ք""`\o1&FJLL B#""F;v8wQFCddQ׮]O???U")hرc۶m޽{&M&Nعsg8[;q*)7naaa?~ܰaooor iLڥ%<[`XN?GK(DwYj՟>00pRc+M6yfv6ӢSDpȐ!mڴ砠 ՗u:M/ET.R]۷o0ݻw;nܸҥKW\!֭[ǏP;;;;וYfׯ_'[qKAp\ko!2k׮kj*RQ"]P ԵC.1uTurTԜhDn?\ 8ݢľO(J<=l5yuJarjcooĉǍװak׮!G>j(e"^ɑOʋ/޽ѣGsAS{进={s  # :zhZZڐ!CPJ+U$:M6*v7hٳg*W';v_y 51_]80RۑQķoveeeUbEjٳK,9rHysU~2C1zJ1˗/wrr\p˗/lݺgϞ*y:|/?Q}Uve)H5ʕ+߿_Ⱌx߿WZ{Q%Ν;زeK^~]ʥyUy ) UT)UT]]]IC$cǎnwI&a*WΝd1 \҇͡x ܊ss>]s_vvv~B;xA˗?u^jݰaQ2:V\9bUR7nlnn`ϟRO IDATReB.$/#^B{)Y(TucǎA0(QƍJOI255e@?2DO$jҥKሼ}u]^:kfrSA pNACYfС~B/7mڄ:Ezѷo_yLJ?^+޾}Yt3gEFF*>VZ9^Zriae/{6`{^bb͛ty7:::"l믓&M6lX`` V ..x4VE&K}Bn6:5k,66s v,+VhԨ#G 9͛;w.H躼ST&wrr_xJ"R L'Q IF|+;_z'6nx3gyTu,o /]4Cn|-Z4qDV9~8)*UJJJ&;;WñzWPÇ 9x L*JlٲѣGQ:F|qΝ?gϞU~ZY߾}377 r^zmٲEe(=&(U.ͪڕP@Ir4zqP!Hf!ufffȼŋCv Ly/q XǏO8puD&"5jT[YY;vLq"Np]x*r+`͚59 _gxZ0yRJnœgaaܞ={(6l߿*)j7u29LJ- B9st[qT!.^ o>ٶm[;;;Hxڵ'77bŊURE 0a 1U. &=\\Gة۷;881I366UVA[b^,Y2n8sA!C :,x~왵o$_~EA':*UJ <|ṙѠ9sFY# Wq< n߾ݱcG͛|ryDm#G,_|`q嬮+ J j22x^J^f֭QVZRV/^DJf]V Nfc={vڴi&&&:uqjB4m>HQF!f]`8,,\r6lXti!';u{n3@*v!;vlݺu ^AтDW<]Z7܊S:;GP;wN8Օ¨PeqzU#̞=Ό3&L3gZT_EGittz۷a!711qƆ 4ppp@`]T)(=HD EH1|pٰQ+UDpNA z;wX"<]"a]W YԴ[.NB\T4Eƍ===v#@?-,- v[n QzQqx!pMJ(;w/^a Zj=K;vܨZ0 հ0sse˖URʤZZZݼy1tΝ?}1=zl۶O>JGbFɟx\7o!W2ʥYU#6patY@DDdCȻ7\( 1I9rBY#711gHԮ]<.4}Ax8Ԭp0 SެYwG% )O׮]S+`!)'Nxyy񛇢ϏxC Yf >-[B ͛Uݻw>>>Pn 8Uao߾۫F Nݻ$6mm.%ADTKxQ 8D&?MyRXuK c%Gæy,A FiҤ -L27V:3{Ĉ+Wd~DDD`` pX\Scƌm†YYz |Ver]J Y7 8q}$]&%%5he>ׯgD8?Ua333yn=:uښ?7C]>vu}vM'"r9_VPhhpPPPpD? NUJQ+e~!Jw[u\W/Q._={0Ǐa %K/\A_a==<

[ݸq n?~Ū:*.nCaU8Pbӧ•^Ўd GU^}޼y2b;, z%URo8jtf͚7x?_|!u5%{< yrTOL||Yc!@p <@ȃߘD9Ҡ|&&&3H8;.e&lgg_1< '>in 'CR+Tp9__߃V{(pY{]"Z>}PN<)J)?~^;ĥ 7xraO9ݘ1\?,A40^8͇v*DSQ!o̞=g <6#' 7ҥ 6ʌ|._^M|Y]4_*&!!χNeN NNIw> A)K8NtiJ|tH!t&{ "oooovQ [nttϟ٨Q#fff0oƅH68,z+D6p8 >z&G K[ݻGFF6۷ԴvܹsO,@ xxvvvaaaPY? ㏨:t@۷/ ٳg󅒃ZɼAhر?'"") ^D6mD}3^Z#Fpk-2B.@#W8+CcOɊ <#!DH7!0nnnJK*Eƨj[!ů⻉+aÆ 1USMBΖSNժU&%%CK,wߍ3199> xe1asex0|OoҤ |M֪UC#n\ަ^SrY5`t???q V\*<<E״={"޽;cmFb[MvڨQTBϛ7QF'N`/0sX2]1Νŋ|U];;;y}C0qxJZ&Nh"#B ,C#w\P!g, nH8={Poa\ AMI  LZǎC`cCCC'L`[[[9ާO~!{z"_111=`+W iy5\^YAy}c<.ѯwF$ݪU+Hرc'Acrssyٲe%JyiΠAHA?|W2ۍsY\K-xrUx5jCV2VCqsYSD `xH{Yz6iŦ^3]tɀu޶mŋ J<ή]O 9۽aKϫ3g0|xɒ%,XW=((I/1/N><bh;: 0dήZjPaÆADrhB"!yY=^GV &h(RK;Dܸ#\WWמ={"d8T{#+N8Y|rss5ER[Ryk¶mʗ/wPo566"~-[+tQDX/]:ZD"ÇOAKNNN۶m333'nŅI7v,!$$=EDy"G++P9iڴ.Ǐ'MȻvڍ5ڽ{wPd+gB_D # %KXti.3<<|2yf `UIa[qܹsٙal\jDGGݻwg9$&&BP\|iv q->}l2^ԯ>}H.d4j9rdժULp{|%h޼cǘ/S.8S\v &0_((pߙ޺ukQ+Ν;U\r ;64|]}8 ;w PbEmڴ $[l ]g~<1\>&={aA/^L_t !'N\~ѣG #*.klO2x\(!aaad""]s˗/ӦM_>)5"]]&{quuussCfQvQreg8 [: mݺuQ#]xf͚&L4q.?yyyC6mP~SXݦ_*̛78)S*Ɔa:o3339UWq} NJ8Ze}{!P2sss7 .ʤ@_̙3LBYJa5OlSNe8 DG*Qɓ=[n7nptt5jiuc5U~{AZ8l#@GصkW56l(1 )99J;y+ 9-- ѕꝑ<I 5J*P[<66N&%%7n…JO'޽=hX6ݻȧ?ٰa%KyKv߿۹|7gE WpK.XI5 ڝ/\% B Tŀ('C$z6& 9ԩ ,f;0~mJ*U|y(7j>Yѣx OU9s֮];h Ui[[[ITVu@ ? Cp|Μ9dNuU:LyeTe˚ZJ !jЛ7oڷoW#>.f-HR[|oS pG5$t2c [%*@FNWeK˗#Bp"_!)|^A?.> x0( %BxY &$!0[Çg%<2ʎVw @dgg ŋű&@w?6mB|)oTY2ɓ(^rj|;#ܹӲeK] —E'%MZ&EA!V\;‘K.E'U EFFB Fe~|۶mWJgmœ'OuAUjv82j ?&ā)ax)p{g sK}~ ӧW [,0U`" ·GjTq!3D Q| 4LP VS%ls9СsA(߿?܅˗#c{D&Yٳi@,š-N]jd ḂUYR;v+牰[W"G6(C -%ݻQ}]=TӧC«Tx!*%Zj={O~EE4i)l;\;jv8K@eNbbS4ͫW:v*eXQFrơʖ8*ijg݋g[hB~Zĸv#MCׯ# ]p9s^7M3YVEBAϞ=YDIbhzϟ[ AăֲeKI'NxQ 0ZǏ0dLX4q^R&H8?y"!GB[[[6s7B|KŋQv IDATGMDlx9RJ۷Ou̙_N7_m |&7D5h֬Y !y-MrnYAVrlH`@|p~bh= Z)V(0cǎ]6ٹ`1ԔuֱZd͛ T0ZjK,Bohh}v\|:Tl`Qo_vufoot3fYTGa( 9K B,Z6box ]qgMݻ7Ty 7nD!!! "g?)hѢU%veBQY 233 K"Æ ])8y)l//䤥#1CO oҦ,Uֳ' *\"!g; R1M4A#ˑ‹Z:TtFǍɓ'3 oȑ#(99m>DR]y.\j[;_z S+W,U1Djz8 {}6`uoz!z1cʕ+W^-[¸tY[[CS+FV]Ň~UwޅCi!666vUX !7oF޹s5j9sʕ+(XjV褑!`Uڙ7o^j՘G8Cf 1"!f2Z1Z/^'''@f=zA [nm@Q p TO4 ~,_\A3@g]t eIvv6Yva#BWS)HG/)Cj*͹s{<:pX6UE\!7222`ʗ/pQnD|˗/=$/BG\ȗ-[6zhΗ~̙3QrPfPڃ SL\CݼyJ*][qpVhh(Y2,, 8p:gM77VZ)lz[JĄ:uItZٳa.5Tz zWŋܢyʥ%ZwP'dr0H*|)**aϞ=CYzu/_F 0y5D333#+++yϟ?߱c+rXjxVk֬YWBVXb"0gpeX&OjXmeGoooDT/8*bɒ%(B-]%T7o@, x-rl2^ . ۶m[dz˖-YfUPǏ(l.\ؽ{7^ N022b?g8sP=٨H*<<78| YVqPy"gx;ŋGZhnٳYSp\=6 EɄ7(!+`Ι 6Ƈhn+5=iMݲց/9p@<f͚a 4Q555Dg^zUfDvY  ~BX6oތ")Up 8y^+& z(޸qt&.aii{|2|;vTzxDNVb|JBB B:u'I@@%Ċ/!9C4owe[% }*UJ(ѹsg{ￓtuܹgj?~kݺuZy9rU+ڈ.; N<5j P K&M OI JK_PCJpMp~FBŋ3f@MLLJ, CC.[, Fl ;7nڋ"B#u0| Ve_~v)ӧOdǏ= ua>tJF 1C͛7m4M-CY-MZEдiSa`ݙB/ryX\]]!NNNrD]jRCUB1b,Z͛74jsNDEb&LIVRcnj#pdƆmX]#Ciϟ?(T^^^ŐT YfAN-7fUҔ< l$H~z*dC&FFF0lmm .;UjU0rʕt(ِޜ Oȇf@0Zax&0ߨPnP_<<z.ujgĉlb =A^%S&M*gϞ_J#!JgX^,Yr(Bp;uqBPV|#"Ǘ <"u++]vN'v…Ġ_J a, CzmHHy__޽C5g8-M޸qBKMG 7YfIHNNo$}a  N?(ۢW[rO!we,oxB~ݻ4h@VrէO;"Dw # +”-[] ^r%]ta~ .a%+U@̄|֭'O1YN9r<5k50VR"pJ)S(*/!wܑ7 > ::oZT\Z!N%$_(%<_Ȓu<={P8Qa(%j/OOOkW\jnjtR XmEȊ_˅wA`[qwџ(E-bرx)D=>p FFBoSK^7Xݻw{`3a!a-!A mpن r#s9COɇO Ni?bh; 'cffrpp@CAmllI2DZZZtڒ9=z`s",L_Ao^z  S [A@Ν)!?`O`ZhҤɓ'ٞӧ%KBb/숑8/^˕%cȤioB/[ݺu+I*U;_zn6M<ʄb޽;2x} nwHP !;R ,!(0dj#&92cx<Džё EClx8R۷Gh8!!/H)\uڕC+8p sՏv AnXjժb8[(jՂ>aB jSa yy vpp77nZjX%6j&a9$0oĔQ gⵢ{xx(NG~QD4b|&-P#kkk\"Q/_";bk/8ǐj6R iD @L:'`H> *Fsgʪg:,, Nġ vUرcYk@ĉ Q!DDaV=v.[*v)!GNA ճgO(0X&aW>}tD([n|Mp)`0CHO4 >M^u- Q |: y(dVp8rNOׯG|7T~9###*^RH.] zr?-4e kB3e *NƲ!#4G9g3;5 <'N:. " ]CH 2v̘1W/GK.ō @ Eg:u* f͚%s}{D0 7ʌ(6ICJIpwhvdd$hHev=`J7vXI 4JR~ Q=6~aÆ W0e˖Ga~A >=CH-_`t䋂٧Bq漥)[*&Z(޽{ K%DɃnժU`` Ʈ]PQ̙׮] w] y%b/ԾϞ=g9<<\Jx8,iM@t.<<9Gv|`hhXV-(*\c|2>>~]C Q{4M2G^w̙ÆqN6k !)_(ej/<'AUtun}$pNYrͰl t:"VB"-)B-\ @QӧO׮]p :%%e˖-8o >{ L"WB|F>III g)UӦMpf͚WO Ve233lRO؊OE0DZRZ5JMرǧB 2eZZեK 6nܸfG8>rqJ*J(&~!镭a{-$*xe„ ( 6[C1\_`X[ŝl!s)SD(HVIr5D#F@PީSCe5 - #""D]B˖-=z|ܹ峜W6pu1O햦nnnˀIvyYt)ѣPtDa8H_ kX́j /ۦM|@|xeYn+2i*8J6$ ^^sbXL*'ݻS“F]R wԨQWFCEHYb:k֬:t(ۻg۽>|ʉqㆴiT➭=mPZ*W,W={ ',ZH4BB#@-M5 9z EN6ùC{5 eÃYUVED۷/}Dt s`nݚLG`'3Sz+s"<߈YQܾ}{׮]9_]<uAzI&$R:K' G(opM%Cx aTI 9f{)ռʆ@&Y9xbZ3ҕ%>he:0ڸ8ggg?rqkaKS*A-Bk.xdePp uୣ^!ٳ',oBɒ%6l߽{WS(hBD$9Ba!h \B͎;S<Űg;E'7DkВiºɓ&&-r9ȷ_]H1RgϞ]p{ｇ@DZikiE\-##úTǽtb 8圷oS&`pp5b}ǩO6187l0a 0(rLnE7D7/DichmFFF󟏟HALL \|>vMR:A*4`ؾ)2իwEnX{9)jΜ9c=nXwCXF>e*‰y[brgji!Yl3fO~"H{Q4M3%%e|GߧPv*KR>M$n ?%D C돡byIivuuI򾴕844za'xFTTp^qhxngh#88X׻- \ڋF饓5U 1׬Ys=E:}}}n mV5*<իWEbhooWgQN[ƶV͝Ƽ*jOc]o Z-]M]=-{vv^XnΜ9YpaUUٳgV@! d{ gB;;;[lFڊU;-"abccigvޑd2*'HNN>{#3ܼyWpG쯅ӧ4mp&^!rBBljf=T(i4vv;S+Wd8m\r۵8MяIpQG~3G.AZjmv]/ؙߗmIwﭳEh˕أvАp?[OwGf~֯_.|;wAbLxb"wYm%&&Z.njjrRJKKcnMJxߣ#mɞhii3>7ݽp?@aH'nq'/]T;s åXxblKk׮ݻW_]Îj3gloo[i1ngD}<] G ]h'|°]YcTnpONTiK,믧iPi 6k 1ZM;22" w=/;O rp˗_rn%.MhΜ&s̙={K_44#(C t:a 򈈈2d i,?}ɷoߖ O͛7O?3ϊ+zݡ}b+u9HG׹E[ CQJR8bN»ij56gΜ1a`&v;0 w29Hf۶m6m ^dlMuI ccv; :th4N~Ia2|¸b… O h4TK8s90#88xbL9w&bMl31yLLLXX|/GANϚ5L?<þVen߾"|& gEȌnz{{.[uO0|n%49(/` uCzbk]͘1nv+4yfffkkTׄ49'md;_תU^uE'9Z֘4Ο?/Kɋ痑!}lGW\d(Xm-bڈ/((p|~yyb5 h4m$ՍU4/^Ѕ411_;=N{gl*-Z)8rOO۵W\aXV &{'h v: {o|Agp<%%͛4ypt]BO>@;c㴴JeGvg̘ARmقri0 9Ewvv޹sG6\9::J?+VVztttxxQvx m ,߫-&gz͚5wޕ s bb`Zn|tcha 3 iʶ 86}iE(skZڄmmmr]kE#O ;V. R\ya%O?_.ذM>+@ ;F.*<1"ol&usΝ5kⶠ~18qlݒ'9L9A}YN###C ǟ]wt[nuŚ=PHH… /\ wAd <(((11ٳRC}] -ُ(bTj{Z-I1bbbz$kc A9ӻKFFFffT% ]9!iNgf\PiiiR )9<"{SXϞ={Vޫ]wHK }rm' V/^,o1RSS+***[ 5::j0"##.L\"O"wYG!Ge?60gΜ/R2#|}}8-/b| `ǏWUU[ 3V[5mܣ>2Xwޝ7oJCCCώ2@v8℅%$$[ZNOO?}P>9L$iE@&ȇ =/NJSTGaEwzC#"#B6Vtd.'TBQQQ4-y1HƏkjIʍɈAG'KuՒLyW#nvO9`JV322 . L{xE ԏ^_6i98XbjAeeerM\4\{ =q(dOϐz_YBK ֓uҒBl)''KZ@MxkzkSؗg 17L+%%͛Q}f(rѭ-<,nzW!U_]n;[on^Ȅaw45>?/wYyJJdh(J"~y}fϓw2 1Jpp4J.Ihgٮ.<nBC(r<69B<5tKUD.XLƤ%.!$,@{] 7~;{z-Oteho($dnWSWOKݻMu:R.gƏr7i斷lf0`wկlj.rAS4yV u߸qoTȒM}̺l8˗RT]]Miݺurĭ,]T_p!)фA9BD8;҄GMy\ ܥG-r$;2I䅅r@xhȊPT&C]Fu>|x֭Ү֭[s̑v.k׮wҮܹsҮ wܸqcUUtz֬Y6++AuܹSu^t)11QuK/t!iY[[+yҮZzɓ'%_'9@p A1zv?_ܣg7ȳZn-r!8 cr!8 ְ w1`J&8iT zzz&UjJAL~ZGSTI.K굱KYƦ ]l|4U4cKաqqqс>Gu=R=>"hu,}\ϰ0ć7T|BMp1 k-#6~?U4} 9j(9xXg!Ԕ FYk'-tWjܦSw[6$Z]4`=!:pH 4fXcyTTҨB޾<5Z%<w;JPS2?Kј ejAIeoA)_lށy'pIȷԔrBef2ܲV2z ݼ ?1yA)ܞ^swsJƛ+5 /S-zg$D7ܠ2^304fp_@5)c cr!8 cr3gluLؘ;جxt rg aomOWt<2}ثqď__?;'$g{Edtz@䭷z/r@~t[.rlO`hhDDOv㣥^x9b",984y=ADv]aqy6_BhZo5ZSUx42{f / _cˎuo/oZur:vB3r7yDK+,ps'7Q_k~_e Ee;Kl8G.L/y87B~{ qvg#D[x>͍NcG+^Mέ'DWz ?!LlVB0y03-3D+>ol,^Xz/ phuА?9r!8 cr!8 cr!8 cr!8 cr!8 cr!8 cr!8 cr!8 cr!8 cr?<|gyIENDB`nF @ P3ܾ$N$ț}w`r| ^(@IQe+ě@j>x$IpWtwyeԡJJr0V)Hs(@IQggH HEkNl Fúd' ݡ'MpGz>NO m wTR*)x <@ @ 1p%KʏJ ^h4KH@ϫh5X'ujdCQT֖ۡ9=,uj@PDhܼ^P*Ȑ:=2R瘥^hڵux xH+kMsWH)ܨ[C[zxHnGYj{cg޽OڏCYA՝iS1Kڇ\Rd xHMJo)˗:9s\5חWTΔiVK xH٭ۛ%JVU*ur<麇˳lj21k˧M@FXإ ;NJ#T Əlne=rN qYWPPR''=ےT]u|HhA=[}xe*Y1ǔw;!uB+zNd }R'xg};kQ6lO)(*..*Q)O '!=ٵ@0T@\TpKmTIx>dѲoMa| T@Lu\yg_|ƈK!f]j~d#mb@FPx~]R>]S.)zbZxaPڴ|>l`e1v]*}.taZx" ?]~'}Cڴ2ty/ <c ]v_!zkӁBwχPLws/KڴE}m*nk G9=lbҿ΁CU9|"{47;Qj-6-i'oVѵwfњ=hu7/H]RHF9oew:ӄ(W(2rr ˨mj]ۺZ>諘b @2kɎUڴVǢ'MҤ/HQ@۞~E`m}4/vwC7W2% ϥ [r%MKZڤRsKݯ$κ!x@JB(jXTh =uNԑ.$ o'M/}X \2hL8}4ڇ]bn){lך+$w֝30怜e|dJ +6@wAx@6_{Q &N4O a=2{rtMЅY/}im2%uZ4ZѦQUq(LMSfag-ږlS$%xV,6U7"Η +(#6Y3&ujL[- ^H)C!xɃӢ4S( =uuI\ =uy\N?S(ǗbZn:_TySUY<;Ol"STN9GM[Vs+y[kVr*IRLWu{qxOFNhv]~,7}#ÆV4qc`+kLJ'^L+o4Lխ2;Uп|v˱Lu{t3Cw窊b[ ^y9JW|Lgq5>iKfMZ~!Rp{jMاҥ ..ȟ {ifm3!+=tw{eָ켂g@X_.y}Y˺|&3Xɟ*V2D'w#v#$*hc^yLK2ɬwNIH"ӄvSm)+S\LG;zk>JdX_KK}nkdoP̷v;F / VvQF)~6*rwζV( 0֣ $R3w39ImaZBZSo;Tim]]zJo3 gttZܵⵅj`&iޑлº·Mhp?e3ƞjll6;Zt;K-1}c;yF }6qTN5}syf _ <=Id$EgnPBh[ʊY??Rbp ? E7hw2#0UMS!e2Dy66 *֖@]f/@}ضP <}*Ṣ4Ƒkw|&lPʆbكwjZSuFzCDan MKbuH\m֥* -M |wRkHOC7vrkL\1֩!!֮s"}R)#~mJ_R^ q:l k6KFerS%HgaS?Z[BI0Zx'ۂL HǿvwCY4g(Znjpώ%A5D瞗,Qο*{8.eIXnxﴭsM Xڞ?nYwGmؚ7plw['SQ8OHsxREoK6fOwwC46yň_^{62vF< :WK"sD}.b[^ivx׎Z6DluLXx%̵?\ iHº%aeSy9؅ ʻX{[}xd~ȍQyZ?:[J;x&O}tσVxT{XL Mr+~EhkKa\6lxߋ`[yi6^ Gֹ5'GOO ȲayV +g&_z)I{z\pB]6_j/y<'˛pfx;bkI$2H`v>;oݦfl33{*z׺!Y6u2,:Ja+ſE>fD)cضP y.5r:߄g!r9 _~YX 6ZI~<.yIXnR*|[wZT,x~\P?GܻAeo$~+O22ofU5Xlp:paPѧr $6J9He~NYԶo)ؕCAս]i YHe/oΜ8p@=}3c2 l eX"}E'ǃr 󇓦236(-CuV;BR{#^D,5Q0;lsB M 3YGR)=#גs 7*mQ1Kΐ?'2̰3^ceagdgaS}zkK`\5eCdl>m7{lhi(f$rks "OU` |]\7m>߾uS O= uB2rrc #(xFOoXNOCj[1<Í;mYsÈX})~b¾K&1ɉxo)K;o*[][Sʒؘ/yw%|YV2. +T~yyo7P+~Zyt3UjMߺl]AnpTܠ?eWIq=۴D4##|5h#`eU_)e·h梤Q*O9ׇH }|cD9CG:F\>ur Pjz{ɡş@)#.As/=vX'vDظmm\JڢKޟ7 n9&2aIMPM7Oz)zyK@~iӇ8)A J{3r9ax{2ڄI'(KU%.TD%| h}uշqnV~9ZywQ U]ob*7:/3?TEy(Er{Ȱ<%. gX\ءƓ̂"?}eU1^5JvrwP?#G] ~{ 8I l'BºQUBE$E;mAsQꎞ >!':l&Li*1 -7]Pn-+ ѣTw6cd$DK$)ky}%p$Dd˒%<򆁛eVǿAE+Ξ-^Ji%#ws uCJ"يp{m3{SzE+7q#ix1|LdRY}"KTb(T@XA.y;6n#EI|??f7ukcVbsS|WIJu*b莲Vbk-XD@"T}ͽj/7k-|pE&CXeڿmkL}|X[Jʆo!pnz0ܦZ5k3@A7Sb7KVdybg%ΰX{oN "ձ@|˗"(uddr2g˝SnR6٧~΢y[(NR@2^~"Sgƫϟ:+RUH3𙛄'5_wz :ofجEfJq QD@2 !h LaghX12kJ)viI LLt;t|J:n9gj.q:(W$Qh4u5fg#JLң*`PօGU) /BER$|P\gg4YIM2t%\OLGcS~KƠ.%wO/˻>{iTkN'֔e3V _}#$@DzOmn2]} >~JWJ.ux:O>41?U__o)ѳwo7&< |5Kggݽ|k|CZfX?y/&\[>yVL1\R\ږd's&?I I/uM۽4 r쥷:/JEMtKeh=/hsUa Kzl5u5*9t" ݍy(-x%y 0Q t{ukMk81p(w˿ay_u{PZymJmpbmu܎>Q ^فWeAHT1]x"n r)ES]Q M٭-@PeC1}ܞZl _ධ{jax@pmnT0YD;fđvJ^C5xG( ?ZJ`161lX8TsKax}Atxx=-{vߒ77s4(E36;>[ee`Zi13|QCf>=v?iۂQegaS};K)U1]MYqmъm0`b\jKPAmm$pF ;Bm.rhTH .tѮ쌲OQm}B5o9l7%BbBEb(E ߲.WgM3<x`ݗ >ݲen6ͭD&;E+(Ƈ4w}'OJ֪Ԥ@>Hbrq&BGxlL4=OqvCx=VkV{$ @;B4ɺ");>߅&m]bȬ ZmU}umn}cMti33~t+к9z)L\Ds ,,_lsbUM %RraJuK-Za^orTg 8,A1IIQA}<v1۵P%_i05b_=83+ O.P1{|gPS |s3bpcUbe0џꎻJ 2CNɱ_r2߸0**e|}ٹ3=ELwPrdnX\zS3Ο]e޸(_&MhS{j>Pۊ7^-$UqfO#q ޟ3wnBcHۊx tyUUC[ S{lkeyƯKot4U1=dU; OL7߶Q]twלz_SUC-gf{>Z»Hg(ի˔.'yo1EDXLR w\*o ((-(wj6a ?^#]IUI9JgK<%񉿏 =J9$>;Rjz}yQ"p1DsmQo?)(qAT=LLbz r·l*ri7W#Yo"g\߰̿;Ȳe"$o+b\z9eCGIDMu7oMJz~l6I~(|G-ͦkdnPD;\/li&ڞ8/ S=m*,[ (:uN=<*0NF@C0\5d*?51?[*2tjx2 N.y6e2efB7oTylZڭݷ\֕6AQDbo}.6&4Tg4۞ڦμHe?w.l[;0ϙ%5L{e x79rӏulk>lC/V[o0Жc' T(=,:%ƫ j[GmP8>=3xHl"^ɔyē0p#Z,N^H+)n JmT&ʤy;C> o)Ksڤ4#wuYL0.c} '<+n ܌2*˔%|2l꿁I숏qvK1js0f_r؉;@Zea67aDꎹicR;^+&FJAiNK67HyIxOkYx{?aWJCSe^.X\-]w_n?V֑xZy\㾧SN#L8xȨie 4TC_%k:(EK9rεzU$^UDD4lQ)YQSr9*!æAnb:q(4{ZEz|09`ω6fM QV'Y,֘A;CIk}F|`GnCvB~haGJdC^2etadM>}vژ$;8MѤzHMdeѫMf vw#h{>⿥BO`#>@b| '<+r <$_˳g4z]46?ga\,x}U&Y}l'-}xzsuouqft ^|)2'#PO&0p4I-nk|Mu>)_L'jANH8-}Sl3ևò,ZMqϒ̤H"ö3lZ㿥tN uP-a5[66sS|Iʰ2Ұ0QEОc~%Te4}'6::Ys% KVlqg+O\ޝꔼIrxƎ#wr)'$%& +|tE-5A>&1cY0Z.83)qRCO;)-eh2kCh/,q~`@魳eA% &U2\ SU-b>vL_zKV2z1xǜR;x ^xbMN`3;u(EzRDB ra17zF)uJ.rQ]rCzDsbZwJqC! +ARY^7SjŖR>xɁ.Vɓ5T/yh$9zIIdSKhp/$)rrss9@ X.[,uTAB?mrqm6i#XolQSX_}6n+M TK.2eQi5p, Y*U(2R- \K236 b'O{l2+{Cmc6I|RQK;ѿUr5+=/(Iȇ|7T,5YR]Wp Jmeq (5 }[N>v_ HuY(^<{Š;gUʃq%yT.1 7q>U)SN eqPx뮓0 WfoHNqC#_=u8l*5=eEuY{ΎE1my=0~n{1{"A <ASCKnQ Wek9 mԇDmO4^{xH7}7~5m_ڱ* ]U*nXlxjQZ^VKdxIV]s{lO>ä+r`ĉy@DSؓKxSӱ!|Qjz>7в=s|^NixOV=?? '_R]pp ݃zbۅoq7f+AD1̼´Hx> cPoW Gy/G>\2Ξx2YN l0W|V]^@ۙϯ5x'w_X@T} ͛j:x?{YR2+]S$kk:7(B>c|&GU'z{k{, <?T>SCH&vo+ }5eg?7?IIK'fveWGp"R̈́騺b.D@xH8']jjC#W7:{t@rdiFOC1v_#U;NJ' 2a]y0pnL?}8 ~d]{>q撐}š? [vrjSQ.*5QeҴo{}6-GoٵwAAfšl|d _Eؠ yE&?xNg Du8ltr?Wd"LS:Yjݥ==*qjRѝ&fA5v}nd\ ^=(Ϲ 6@)>228@nl [(~1FսHCS,j bw99761 :Zjˈm>??D7sAGD]T xL$}Zc"`)^Qܪ׭viFFl5 Kɡ*,*d&~D(OKGD.*@iA/|g4YҰu~$p#d2%ÆjFu?N/(:9 2.AsZJH3aCT !|5K˻>sQ^.GǓBo0;swc/K3|jW+!].99$b`AvކnJgs>~mjÖc=Ǚ''gaB' lo^n Z==]ç(|5wW̎.V04yX~ #e,7\TJ y Ul8HT^*CLZm8huj>ݕP}mi dPQdbwyOҵ[ݶ"romt|D\ɵ+U36n!D܁REyv}`@z˺@uȽ[فg ޣK5幒,b:F\ZoZ$D_Psr P{[^B S9^lZ V Vmm [\z-ݷucۍoqK!'+ Y'ǘS{ Qi[=J㚁m>O3K/ &csfd f9%gB+Sywqi_ll&DdihE̮gW @c~]JmY ?J|^6uX4x|_,KJ{%aێM ]!rV~>i;*Q(pEZן}h[ U) tIe>1+ KýĘ8޼f܉ )nX1ȊLxPl~];GN~D.tMZM6L` Iܗ̈SXCgTL;}9Yh7=AЄ PHTR*)x <@ @ PHTR*):ANIENDB`nL- _rAX(x 1!PNG  IHDR@xgAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME eT IDATx \LJJR}H,E]]rvE {Y~fKJ).]YY EQn3M8sNgf993X+S ݐ ݐ ݐح\ӥK!tD+rvSru(NAeM'Nh={6H:O͚5qNӼ ɫ|-\2\n\n͛heee@H @ H8(UUU 8-666G>sL^kcDZboj +:FsyܹSGGgܸq$:v8!Mttp[gΝ <Ϝ98Mr{W:ujDDٳ<[u}||; kkxnYYBh76rFYSy%%jBo.ﰑ#Gٕn޼t Ηs/|5.dVɅZFFFZ*#K&cccǎ_^.¦\^rJRE<@+-}Lf&f2'Ο?o-]ǓTM(sf_ܚp2k[jVC9Opk$ٳgKKKg͚US#|꺋 {-=^-ο0.f=n=ș8ZlҤILEE%224cbb;P@ǰ(`Ѫ()n0Q3G:ccf7WʠwU[.ή:|[AA! &66w+X˛փ|ʪv8 I ૯љ?~LLZ&|ƍC`֭[駟ӎMsyv9E9Ηr9!SUUuȐ!튴-vڥK U9I_::n\m%=mnnZfwڅ<^S_\OWTTm0Q+Mf3To8uV㴱8j#Fٕlٲv8&O5,ˉV3 h4Sw}ѿ;At4.zKUT WR[^R5ŕj:ޓTH|ʔ)K.SP;@{Q9SPU~-Wxr3/K2i۽R([ԣ+ ]W讍43gN׮]CCCUTc9:_Meޫk|>q%$\da(6r3B\MVZtF|ߑ=H;-wlRN roQg^ Jޝɉ:0Uܠo޻0!{1 #=kf&ΡLAAczzz#%\νcs#W5{ VU]_\kob[kx[<#Ǻ{nK/۷%K.{ݹsqΝ;״@SŇWi*]i`7^JM^S򎍙a 5SQz}[I{g8fOVׯ_K/I۩-[y1bpd!{۷oH>>] (copB".SQQn`8mWN7]dVF6On䰃qۛ2 +7ٳ ,8z)Shb<O__?))iժU;(|5'gt6遌:0.ϟE0xHYj;… {EGG7LL>ё @ɋnG]xddefxhZeVҎGu?K%]~1e˖{ҎOĻK&w~ 4+//@HlJ_WRo+mm|Z}J;6 4iKw\S}٘{ꕘ8}))) _&00pկ߰a6Ént scGTSU_Kr23 m5ጷ(o~쯫suu}ɓ'UUY<dܽ{wĈoi722rʕǏ'I]!4y9~ wu"lR.'>}\x1664i vuΝC̙s3fH,0鞧M507e,SӦe5IV5\ي]}ٸqΞ=ˮ1p$, $֗0aB\\ɓظqh im7(I=wέ2;;]9h8uPM!ڵk{%caaA;xYy[ 4(==󒒒oFv9/mlQ>u,X}׀:uԂ h FTrJg/f7~OErY9H ` ~LE$s0ֺqy1R[[Ae,,,nq髪.\(xڣG~iڴid~chQF[N0]]]}Of={ӧ @$~WWjyOqdޮQ@y:ؔFYVVyܲnoF[ 0aB17qtp+..vZG1""JWWnTCQ+m=x8Dy5jF2L*.\~)R/Ǐ{n'h`x<㶂|.ʁlMҖ_iK'[oN2ݶm>}_Zr%iH>0C$y ~3f @$~Zf ^`yu`W4&#Ũocfiz n;r h}t̏?X[[4~m rڵkdRCC<=vIh.4dw7UWEט+xiɻ+nW2Б_Ԁ{tvi ݻ455BBB&N(@ސE_5ݐ$L\6–cv>:ƪ:m\gM:]ZZ*tiӦCf͚%@<{$3gΜ}x6.?rȔ{{{ҔߴiXcyJu$mtPWhD==Ç}mذ㼂Ԅ >|PF!C~Yy[yzzex<ܘD **j%xzU\V ڧ:""%.Ν+..={!hzӧ-x|۶m;v ]OZ'O;vl޽;V iK-Y>ۺuuSZZzY4,,<鉶7o@q./I 8kPXSS.C]@uiʕ}9|p{W &L(**LQ H6mZ=O/_,&˖-{hK1IE^Mws,ooI&Th:ὢ>׳\+V&V[}W }e~RR҈#]{^^ސ!Cn*jkk544޽K18mm3gP|Mk]y.v( zWp 6]vy+lٲyfkk|ɴ]cbb>çO:`QFӇb $%KH.ϨW^{Uzڑ䚟Sē{8I-‰'H6%$%Hv.⢣Cv $P#̙3cٲezjڱr4u^Va,)G9xdVGh;w܂ ~ǩS?~ =,\*<.t-6א"wILLQNN@)̚5ҎEv%KЎB*Ps0}wFx2Y;¯} G1g|FfjnkϜV?nY'߃L"yʊb֭۱cDŽ Oau .Ow~}N4v8իHO.dU~ռuHjvk#2 ɾ#GFQ"-vpݻG;2.[RPE[LԷK}j=tc%]E L6-<<|֬Y?0Z IDATyHoFZb׮]:::5?dճ&mllhtryfd}oҞ"#GLII/--ݴip@BCBBXabҥZZZǏp8a۵kh}AJAkc .o:Dك),,tpp ?|`lܳgOrr2@g>|̙3i&$oݺݝvD +fВs)*Scӕ㉆Sg HH ::m\ݻ>iO\;{7e2 >/ֺՅ/MCCܹs(+㴊zQ\\iHGٕzzx{Hhntѹ&$[A"wrmK]P𜊊JTTKhhcϞ=iALOOOd㜦XZ˯'đGV̜H9Qԙ/1q 'D26QPP F)**zD__Җ/_HcǎM>YTryyJxʱgk_dwc~FTJ]tٷoߜ9sƍwဈ]zٙv qgmm-pW]O E/Wc0Y YN.%䱇xGۿG}$ֺ@2>3qqq&&& ?>E5O:aÆ w^wgҤI%%%կ*͢.Wrrg$ysGN'TnNjdzXބX+I3qtڱȐVڒ%KhG!2 AAA]v).._x1yyy_tb5-vtoۇ[>}4@@222 D; رco߾M;)SCCC 9IR rMVYOf;w|rrrh ;w|oPSS#2UUUÑ7o޺u]kC9;;71&ޜ#r(kGQoK}N}n 3(eGݻ h-z7SSSiGA[|1ckN۶m[vmW۷իWE\b0Ԕi.PQ cњ(8ø\Cf gfdW14lH>dɓCCC9pU?Mݺu۴iu|||+WjGՋ|cbb,,,DHlBL24^62׀6:nxx8@U[nسgQNN@%--m/ E]I;q*1~=,Oc¿$c:ؓ*#Wģ8 LMҮ7Ф.HѱlŊÁڣGQÇ/^\_߾l"&/ɍSS^ӎ&YՅ0YsUyϪ gAB:vUJEEԩSѱv8r ؅4hPjj*i&JvՅ/~={<{,3f ;BHpܷSʊͽd'U+m/M"M #G0aBQQpNaa \\Z ) -[twݔ#""IR_ERBc*NAdf5푇v94K.>>>g7n۷i#__455ϝ;w}iϝ;w],ZѣG---Zҹ*Đݦ}Rc իWю}K#G\tI%Ėg|fne5a}}uAnm .!!ځȾO>dʹ`;v 8۴i+I.|aaa}yWyBWgm\txIfnju vE0HPPPWWזypBڱ,Ҹ$_~eذaca_}U\\[oE;>#]]ݭ[Jj1[`ANNf{n%ɫ|-&r^%+U-lB]zLJv C]_<={9ݐd?N;!x<;wh@dV&4_SP g?x9*j:mӂ%VSRR&MTRR_GGU_y\V.'mMtrrNΜ93tPL+եUH)j H2W~Hs~ӧOiz$ِ`:""D!::zɒ%ǎKa׭kWjO|d-.U$ŊtttOI"vcƌ!ԍ7~ciFEEiϟ?v 2MP]]ݔ)ShV8cƌƆ.+7|355gcy544h"ڪ{pĉeeeaXm3g I?lTTԪUvGs[MЎX[n$[YYݽ{v8,]PP@z$ЍJ&$''_zuѢERrV|)Sh"(c.44vn^^^䷒󄄄!C5I;itCajjj ,:ujxxx=CSRR\k㏵'Lo={F~ɟKDžtmݻw {&Mti---]7___6Y:[f`0+K;nnnGqrrs󚚚AAAM &ݺup8ݣŋkjjƍG+';_WǀQo0mڴ ?Vra1yJv(#Wvڵb J_N%i\1[܎4Lb'$$8;;\v8RN8qUӇĄnTre666111撬ɓ'iiiTD.>p>&$⬵[GK" ȑ#SRRI:߸q#p#FXuN<-**$V;,YDSS1,,$uollӔa\e0Kfe_)(( qÆ #Gbcc%Viӎ=*&K*+[8%'';""Bb5~u7|Pځ %{^av2+?4Srp*&:Lkmff&8A>>Ptd_eQ _999}W˗/w]\.O>nիj#??ci+iHy zr82oV2AUUU9::TWWӎyĕ+Wȏ]-Z$Z]nܸѿݻw~|]K{tnllv m%4y _ok4wܵs=S~BoFtt4ivOEEp8[o-{bN)4hР4q HOO [n3`1r [ͪ4i! ?L 3f͚x  e9ŎǏ߷o "ZG w4t['O>| ]w]wޣGNڿQAA[B/hr|?{c޹sv8/:1dW#99ҥKb4\]]ȥsH%՚4gLDϞ=>|8}jtL˅XXX4ιpp@:to\\\K='姶v޼y$"7r!(PGi=GFFdSSӃ?^e[bEnnnttޏUh\syFFFѣG Bhhիɯ$iӍ_~w!_} E~zLLL t@ZZ{4>XQ믿H{8w#Z_ٳdq"gftcGڻAAA...gΜ isO4<ݭ[~ʊ DRt{בTߓK>:+h"OL߳g… ' ttxK9| %~7!Ib '<<\bw߿7:Y]mm-o:YHҁ铕r>>Tj"- ŋĉ)OOOL! IVGGy%>|xZZm&ڃΝԖ/_?8::&''wF\VR?zKlll aYfx)ڱv+-5A;";%%%vz$IkDmmTWVVѣ @pVd[UUճgϢ"G%1񇞞#sBT_/&~m㴃H@옔4}LK(J/^XSS3nܸvpNKkݐggg^4f3TZZjժvaee%U$oM퓝o'פ[TWW`| GȏI;v7o޴[z ھ)߲e˔)S:P#HI&-Zhm_@G>bKBݻG>3gܹs'Hp}KAz}RRRܱ/ivݖo߾miiYXX(bB[q|||vÇ/((PQQ@xggg33C /Rݺup8$Yw8dW --+V {BXp!lINNrJG=v،3d#zJLL$ͫNhڵkIsu1U!ŚgϞ{ܹskkk[_ͭ5PSS6mZUUU {KzCZƤri>^2A./Nߣ{|VnvF+?dV7,r9BAA>}ڙrF$K~#~𡺺zyyXk!8+A?~QXXH;(IKK9 M^{BÄ7&`Kl*:H__?%%e˗/YYY$SOo>.kooO~Z#HJHHpvv~$gܸq \A5k- ʇcڀvvv-tYц1$#;wn֬YW޳gr{KXwyGKp@$p}$$55Upޖ-[(F"ScJ;#̭e^5Yk x˖-uA⤤C/rrrn"#ANɓ')rJQqq1ٗp8d[LyG;NVߓ&MWG}x|7m5WTUUI>1 efcccځ@󪪪o7ߴwE}kƟMe'cRRN&7ޠ5NbE" #[jܹ!/_Nԏ?H;ѣG$\ K ݐMs&(_XX(2׭[w^іo;-022y"wŦ>ʐ!C_.2nWelZ/D[7,2KaE~_qFі)&cǎ_E[}֮]+2߿'2 #xoŋ@d1m<c%8`7%_rكv9!r9!r9!rP5%%%Àf\ޫ3ɤLMy~~~aak [J%d34ErykʯNkeגŽָL9dҦR-###5E[󱥤Ljd3-2bmZa`>;?=Y-%m]4MJ@Sc-*q|jMOGN0Qi7qHHC}7\l2i֚5\>yc걥[?+rw"S!/o?Muè(4M&E LJ;a7xh>NgP++4]eTTN孨`Lh! 64meG3Kc–J F%{{2O{aKryx:WaLь ZM&YxgU'2l)iSxB҅[<w:El.[ <0z`&6F;8!eٛ(;[Jp#\-m-O+?1RAliMF^Af62e-cH,UiMI&2Yd+xG gcKI5--$ +ߩaܳb ֔e@xT$cI6) ti\#X שv*dc}eK&c l))-),oC_r솾o\n\n\n& n]=!JښuJ֣1'@.F{hy7WM.ӔTL FaKy\hzkK*Zoмچx/fl2n]F,ڛJn] EQMKKWWKv{i~Ȳ]l\1 fkIL ҰOb=ofL]\]7q8-a\\nc"Uo\kdd݅ۨ{/P|72LQ%b4#"3"<9 3\ ###"oG%>${]:pdA7%n5@; @gp6noHʃ .ޟpn;<׻\fR[ Vӷzߟs󨤛"}`jckφFDTJ.+%\A.NֆhbUj .p,r;-Un8JڥR%F8_bg`nNv{'>o Z;Z^ : ڌpnuzQ0t q'Qdn`0Sqt-)vVUj+лij{Rwz PwZ9wԎUn] $8!^.PNG  IHDR{0gAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME $"4C IDATxTם{pBNqJ{ 8n6mpúl#1+Mq m[ m ^ӭ6!ЗTV4H9$#HB?i4s̍t5i~s7w~M~60 5` hX0 5` hX \gv~l?P^4\\qډټwrDEm! b$|,f/wx \Dj "@BhGJrO}mwnFާ[i"nk}La b$|!ok 2|oAhڭ1 )a'H ZM&9 a4TLFa4A k6mtxXR3{<래^g\o0QZ{@Ikg%w$ WnxĶ^at v3PPkbl5wD& pt_Y@$ p3t 7u)$/޵xY%G4 Hz۱ek8STA56\%A8f&o.m3fC,*9 Id5hߍ҃nC03qû6mٺ)۱Z2ԷbN 8Ѿ3GzrQvy?ywHkH1lE^ّ+c!ƿ8'9 U 2(rF)ʚϝ;ı"& k%~`=R&}H-,G]ן2w+Sa)8IG;Xў+_C'"MIJ jQRé+ b[EyhU&ѽVRq:X;>U;asw (;=Aң_ȶaF/Of5dq JOk]3ؐ `F;.idlM9YEwȶAaFfj_+1ױ @$2x.u5/J$W,7D6ekiф>Eؗe:g^s[DM7+/ta^0 dףl-iJ?)'N\ˉ vNȴeo.ۥ.P0ؒ4.^]KښlJBG ZjVzwC#[[mYYYxo } 0H(lIFnC ~#|2bLM426+;bQl\r/6q;uuztDnQH&2H(lI28] 5*%U8~- +N|oDXFUVM-{,S*ytg/&@@`E0ʋtP'` /ۀPtI i;EW\Qrni3* `>v_ h[z]"}*g\mB|{.ɰ:Pdhdai@tJF쁽60\x@uCt{iD$L] Fwn=lXL$ , Yf"v%1M |,6˂^J۬0?Lg|ߢ$a0J{~q_ӝύ7W71&-nյy)}3[~ |dE>`D2|7.o} {ƛV޲&n0 ST}n$+m6W; eбWhf叇w^3bqǭaQΦr\wmt)pZS^#.EY1u/Мy2_;~+_x$=$#mET?IjUDò 'I|2 #"Yc  ,Iv7I3r]Y<&}qɿku/,gB )(ae~ ܈Lv߻ϛ&4ޕp&"IFHV> /&!+u^8 N-\8kȽ'&T2bD IM*X4V3"7MLue]6a;MONNުٸr$B3THTŠ1%{S9^ixc:VTQ'Q1X\6 +CʗÊ)6Z(j4=: 9`H[$h.?%Lt(t%7eg^ҮFZ'hWC Ox'nߜzAϿ#1+S**q]Qg㟩QNt תOQPZ'9g2ex3Yа4\ 'ڔ0*l1~^Fj1A^G}wq:wJ1Õ7)׿xl{ǍXjk$;tO|>/b_݂δWz2<87nj%zO8<Ɨe}_!FFhh$ t,B(X4W3"M튛q{p[ 8nwMhۄ9`Ce93 浫tT[xw Do3MDݎ,o/X~EjIţ NT>Kv<·fDёq'*3gZ3s_t:(IjpGل'j:E(X4W3"MF~z(؁0Z`tW)fC3{s̤CF tо f9}!xf˳<{}`0LOO-rߎKn |D+\^AqQEm\zZPTϤKhG8fyμ .Oʖ$a5t!a'ðd%3HYÑ :&vq"ۀ G@r|~C ^%'l.P8RR$K$yc dxbbcLwijKFufK6]_,C7=^Lش8 ^LdXMYcy!̟҆#E5,hcjsntη!edžLɫppwDN(s,\~ar8e/וzٱNBk{z tGn:|G쮤᰼B?l_ȋ/mHmc8H;1Y/(.))b7mZ^ҏu_>DOLLS>%E΅YMj:Q0" GjF)hc׮ѷKq :Ma7nq.2E't pZn Q6e8ߜ>8yZ Tk62UA.͇#Psl-5I%peVM٦cz!̟#u5Y2F.򆓥s{i~^{RKP.1g5P- o{7C;,=2_otVFk8Rx؏9|E%B013 I2EȻnAǰ'H}Ñiѱtkh1HAST2եmEغ~:Y97̙/.sp-@WW ?+ OUd ~k"JفB),2Vt}|4q5~BU_[8ضQbȴTJJ(Y^ z\AV񱕩/ij8jv;!Ɏİ%Dl#x4VjF4/-Pt,֮a뭟;Ļ#n;v7*u,5"&jí0lV3?+A`I2:Z!5&" hPrg>є2`X׿$=$K /6o(W߼[ebL*M!wTQMCiDR]cMr,C!? Y&_Ij[#+u,U^Uk8V3:&vWTHlϕG|(fbɶBg8xf8HvÔ+W& 3sbZD9?y%5 ШҊUe5QByd} "AH*_z9[do˫:.M V2\<g6^> xnG,^7LbR]Q!I t)p6[ɪP<29=+ clm++DVibƑ_f`u_kWؗtү/~ѷbu[̈́O8<}\HG,jһq̤9{3E˚ o'޳" kc-inߺ!;Mm+a mW{ߨAMs"M3?Ó_ܔз8( EwEUg`3J W$@Su Zӆ1ʣW^Iʪvh%v*jvsF0i(z^jȘJ1vGXo. Ef'\/'ʘ8נH2a^M ?ERf5\P^[,N4 Ff~Z;31?Kw5r2~zڨ ]~IUl'PſDk 40S*h_9x4Mh*yXz LnLy[}mb 5GEiiޖ(4j82aroؒH$KSB֛T@B(.n&ۧķO %F L&J0^si8m4V3rA 쑽AScL^mt^<6)*:4dyca&c 9g16FmFaDm O8C0:jܖbC>mLJCݢ=4xui|۴Giwb;`#ދ$/62oȮJ9a=  'FKR8kLkM|zV SaR]2l3$`kal}2;l)t!pM8)%bɬMvyѣz>uet-t"PJ(Li_ǸymIYyGOcݭ*N};eW7Tm˫s{6R ;LHlQ }Yo=Qr\v@ߡ %{9HtY~A2s%ly%V>`l-_n?xgۛ?oCzl&Z˙t}5.P0$a"FꦾT둈pM*=ohh XqiN,y?bj>w߿ʲ ]`!I]PJ u_9jx<N$x^$40JCO఼J?z/ߍ{FK 3 R/Jt ]`$YA^BCRҺbM=`t51Z:1@{'Ѿ_~pˢL,՗H(lI{_>Lu\&F:1aضr(o<вzl)d^^g|D~H(lIZ *jd;ѿvD=.Q$F痔 ,^!=.S+Hdm-@  v:%|Dܥn_I&M2.stk_+)@@Dfz%]׼I>T-9wpƏOѓG+ {`FUgz@zp@wJ+\s}ҥgs2lYM'qt߱ ldyf@V*巃 \\(jʯ`0nr ;in4MQuXp;C78ّܸj ʱ4dx l8scZ5/ý>iӦ͛l)q}bE} zaFD~L`ؑE>ˆc}TA1omk*YwvUu!bwlm=,fVuw`[E'QT5Π`-lhj猽Tܩ |@7s1}X67af!s6++UQ|a3/%z{at0\F^=aU=IsioU /lvwsϜ%˯/-' kejO3:I}qX 0:@u%:IЕIECVjB|seFKJ sL $0Ar<}9˾mL3QV ΏQ lG54f/ֵ ީsiH&_w{+KɯQgnگʲvxh2]EE8ĂO6RI˿Y= 8Yn\Z&8ws; 0 d_}xb?=DfE/[G A8dv6;:3 "B$> z1#D^YO'>=/wjL, ;jKܚ {-q(lTφOFj~ŧGI~i3j0:*1B$> z?+i|˯[stmnnXue g5[ :kW5_~uyk]nGC| T̷!Fb47$> z!H[PYLL"?;t*/OSgkݼ qLw%!A㕱Tv .W!֥%;G%> z#RP!Gv4G~=}8Y~9= Wս|1Gd> 8$aqu#{dՏ7fnUm8ѯ ~tGe1=N褎X}ϗ˃'k}_^pv wE"\歟^ +{,Y}5]Suc1;9׋0~j)rLm߆ xPbtpq}M,@@FFK髆Dgh;-\k?~e9:B3Efv/$Tˉ 0 gE"ђe5MmMZYAIhtf< T8|{ۂ]7|LL\OX{sr 7;fOh/p*q)=@-UΝ  CS2jwVX|!|Ć:[KR.KGb,G 1d(16#D1ĂAN W,5ַp@U0؃:U5-:OǩѳG7E_TRREϛ~xoĊ xe[Cr NW2̎b} znX2wLr[=+MǾ E`/q`F` a4A kFa4A k?H,AeIENDB`nrb_⫢`p9*x|5PNG  IHDR>I<gAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME ^ IDATxy@LURR*d) e)\g F{e7.ٲfz/Yʕ-KPQѢRnos;Zi3~3g3|wVXXȡP( "D BP(j9BP( r BPd B6T) BmS( "P-P( EZNP(lSJSZ5&P( "  B6 -+WZBP(J{S*cRLP(Y\) BmS( "P-P( EZNP(lCBP(نj9BP( r BPdJ!!!P(b*v k'''P(۷]ttiC(ب9:::R@711IBPyߪWSP9rҥKkԨA ",Tj9*ڪU3gXZZBT?@R5jTÆ ׭[G R1B }vLL iC(y9"|e˖]v%m B02} ++WIEE[ 8 jvZNL4F[l!mB^<7zs:]?[N:TSToLtr .^! -yyg+mvcЊ 4P*ڻi}`زo[S}_i<1 EdggO0aΝ***mP(l!س֙μ&*ڝG8s8Lg_OH>zY9}C¬XiӦ#F mBR8&q&y責{c#Eb:4)4K; ITbbblmmW[[- E<P>3écwSr 8P-7_~pwwwqq!m B啄j9ErJhh(iC(8ZT)rL\\ս{7nL "Nr3n8 ӧ6BYÕu֭[{=gP(b$22r֬YGP䎪IIIMPLJJ۷o۴iC By9"\zk׮?$33FP(1B[>"@˻wN B<w֍ ;"󤥥W^- E+^4/Pd$VVVT)Ǥ 7b7 EΠ)B^^֭[_zu9Ҷ4/PdY݋!m\AS(MFFNzz"i[(rX|yvvݻw>L>/Pk׮uԉ 9E&`!/.]z9is䇪/Z'g"Ν[Nŋ6B)|55$UUU$艉v"ml@Ri[(BSd;wCɓo޼!mP*u&++aÆ5k$m R>>>߿_~=ϛ6m"kL ]ZGᇲ 1qJ%~zSd{ʔ)si޼ŋ4h@*<Ѽ.Jz`BKBrpի/_!J9|UMM-11Q]]ݽVZV"hL@8/\VmuTiRNTW.[Jܿ_9ܞm۶7odaru#4DiYRɉҥ iC()u===GGG///"V ̈́ 9"~n޼iffL |p1M6eeeI$yU-WǞ9x{6KP(t4EV]Vjum֭֬[%owwwk:MEF}v֭SdW?Zh7Gcjy։/"e҉6f&fx *w~i:S(r }XN!sαl޾}{SSS6n^^Chl(b)ub)yj9EVx/\pŊlZD s,(vBZA֮j\ׯښBB*w2Ç=,"{9l܆1uѾ[tｕi9)JIׯO |J]RE?ӂ 9X5u8ggz 7wpz%=\b٥@}ׯ mR>;/_\Q ;fiiɞa&%%855UwԬz1æϟOUI 9"ЇY!111??s)((̛7ORnݺ7t;=1Sq+}2#˙fԨQ-'b "h걱ZZZmP!((ܹsb^^^ƍ/^Ro憘f֬Y}| 28M 9ɯSr8E]j7m|^QhdOd.iEvy҆P(#=5k+V8|ح".ŤI;> yy֝6u,c8Q6ۚ /c?NRa]oB5kvԩ֭[l}}۷o1{y=4ﱓTms*p بJ:i;::B)?jJ׮]ⵊHʻv<,?//*PPQ*1d(м"hhh,.M6-^IDHKK{e.[Y[t,o+tYe# yV[%ʗ]:t(#:\"P- %lݺuD"b+++sXܐ5EBnW;FMNt;x%C' /&jL$"Q(bZdVP(S!C "e$ò|)Î_k5Mmsz~fN#pT91ar=Ԕx653P+Sv{(gC˗]oJjwۺuk˖-IBaRo]vС7&b)K <>$/1U GniӺ DߦypSr)yT(!4!m-J>öm-ZD$"|ŋV\iKSXϝk:OR<™Ùd:zBrMUN<,s΅ $cOe|"dddZZZo\zȑ#HYE]t|(kٳg#ٽxbݺu%iht9 i[(g޼y^^^- ZEoڴ yEbQ1V(9ksj2WJO-PBiJN>Ν;ի/_ !32J{]v=}IDӳg,,,$S$fe)4}=y;t,gDͳKP( ۂ Cի'aDm۶ m=|SN-u ׮]uPTT,Wqr?vNyuQCQBrPVǷŋ;wҥK꒷J4"""I[Aa7nkNI`㤤y摵Y'\:|U9̔&J9Rr˗@ׯO0@ҵkWVPآXܹrI&5hЀIDjy֥E =>0aaJ۹ZE&^#|BxnݺyС+Whj_~y&mwr ?$k233>}Zl`\SoE#{ef&i6lƽ'}4ÄHpj$E@"[Nիك2DGG7nXMސOݻ붽bŊ)Shhhh;vQF ppb>rQick9\M%NZE޾HrPd:v@d?ZZZ  \u떩i8^ ~i ﱫ@2j#ɹzN>i@P,oo{{{PLB`UCLISAL)>1)!ŊaH)qM xoe  ]Ë֭[y䏞={={/iii-"Cfffڵsssxrŋ~*c2üZhp}1ΐP7 E=z.^oݺuڵk֬O'uSoݺwߧO^e;FDDDXXXKP'*5VSg*]tSR 8 u5'Jϖ}v:\>,oܹcll)-"sx^WӶGxIG)omC%Mm}mm*`VX1qĆ ܹYi*r7;gΜ)3 l E3Wrexע'rs5ΛHS(%ý={:o޼۷o.׮] m- `SLA=w˗/IC옘Yl%%/L#2ZI6?Qdɓ'kjj"56mZ5H[TYRRR7oδAÇuF=z_mݺ H[DP\D6=,uҴ I|zAOHDOMΗwϯќ&J ?~ŋqqqG_$f >}! Q4(ZR/qd%U=v-PJlٲ3f?~޼y1B>,c ݛ?>r* .r]TTGvj0 '|u=oDCBÇL1>>}LLѣGI[$oܸVX~cvvvXXخ]HCϟ??xs`RȌ6+0^[xI$qJ%1bϿۮ]`戇,ee弼<:Wܐ0xkג60/^fsDz?cG>X+&)6ٔYv"*)W\qww?paH[$n޼پ}{I3A~~~ JII_HCR9lyzi9ꯍv*OΟ&SL'/a_Y:"c,Yӳzϗ<4Ǐ7mڴp©S^VZ-" U$V͊v01kV&zS$MQQQ_ܠ:u"ȍh7yeNNN#F marrrЊyKJvMGτ vG@ʝsĻ:N/F~nݺ۷o6GltHB3H7o|ʕ>}\x9A,޶m[R7'Xr{rNTyo3a6'ydi;ֶT UP'Nxxx;iӦ-T=޾}aڴiC|Xa}L04YE86^;qrl AN -ZdIBBѣG={FqB%[lٷoKll,isW.\Tl3s먩`dj3S}5:gE.ڽx>qrRdDDI[A3̓+Wz{{Y{|F߷䄸 /^' {R~$lbl E~ʼERtRhǏMF6[>"7!)711M89R[LMMk׮M>M-c8Q: /c?̇$~彆W%`j4l0%[S.]zիlvlliC(I=P2هV\Aަ>G r){edTBPJv:tIe<vB05k" segg8Xn^v|? ɓ'm޾}C G|qfI4iݻwr|"99Ǐ-[$mHq82ɺu,,,޼yο=**jǎ4===RٳggϞ}a;X%3G_Q;e\r {0`^￭[>wԩSgϞ-K~Dp4Du"RLj}E.\DzSNoߞmxѕ+W6oܻw@\bJ$*̴~9EZZZ"444piѢٳg'Kfff7o,#Ry9~M*AV>咘x Ti9A&աCKWP% a#??)h=G4 r|QKKKN ~ƌx1jԨ'"G^]cCFFFZ:;RZjnn.r՝;w455qI ZE\FǫWhkk糕`On P-H-SL={@+y@dсP,ǎ_eccS֧|С[l|rb̖4h̙3d #K>}n:nܸ mnݎ=RY<)dFq7t7x𡆆Fjj*rƍÑknΜ9p^˗!?}A.]npE)-OII[O$\0w\55'N0o$nE*֯_?~x ˗/ׯ_իt}ѣG^yC{%TtBDq]\\$\ꪬܵkWϑ8pUA GD޽{۷3[% a8ZD'uf=^q_ -RvL6K/rd!a(@DD˗/KZkhhX{ƌXB:dȐ'm޼Y\#!PfR!͚5+/E5k|)^'&&ܹQ 566>~8jׯ_yۗ,Y2m46J$I7~s(}"- gR)ϟ/Ӹ8-!~1]߽{wر]vh.]̝;Wdeׯ| y-G,Q>x}5&O\Cʋ|%RSSŊL*-߫.=˩__9tP%oZ&"QO>>**n۶m[l ErFUEEE++ ^x:mggWnpϟ??rr =F nŏJ cΜ9իW_vwlmmq!maΜ9ҬY޽{ԨQ=z۷YbɌ bJ1)L )Vt;I)Kؘ֭[x 7ܳgJG;+a*JDÇVZy{{<mڴAR;\|D+^쮭b<&3>ީE-栈b瀈qt75gϞ(4UVׯWF࠮b ;zhAD!!!rByI =#}jjj+Qxq&M&MԫW/GGDž 6<Rk֬y]z~~k$m$JJJJ48˔rn"\u;w.3R H?z#2fsss;;/_>|m۶-wlV\? .rӦMI[W^%m,&+KKKy]2 o޼ݻׯ)k7ÇrŋEaLv//O$s|0ό}`XZ27~i+J Epp0oKNN΢E455كYYY_.rcbbIQP0BٺuI[7n022"m,怌vڏ?rss#mwޮ]::: ӧOKNNcѤ<-ؓWHަESgW8O^>A^EE4 +k=qm!wNMMurr_VMbK,7nlذ#޽{lD9f >XRzy-(%5*00b͚5:vXf͛7o:88L4E—/_ 6mڄ%77WSLY|K'a^\c2BB\7R6I^ΌYwwzPD?Ǧh[𩖳rɒ%B|ܝ>|W#gv$<-†[Oh< ᪷i(3Ҽ0ǔI+<'(w:# ^`/ޙ1E7JO˱lD<*r_]]R{uY@2I3g`؛7olق~}zaĴEZZZfff>;v_~gK\j V~CBf@jժ+VfddHZ@&m^tiNe-όo&u8Emo=c|FhD:cJ O!޻Oeŋϛ7OOOsN}ĻQLL :4hРT볳 rH[/B65kVݺuݻqIII۷o?x ioԨS:tG2\R(W*wR=ކ׹9ojѿ @ 968ťz/s<:3ǜ1mV<NB])BpȑaÆ G[mƶIH)uFEE;vҲ>999SP[CNC̚5koQHH͛o߾]V-:u*bq- H,1w4e޽IAƎ 9GF{yҦgȐ!***B"j(1khhܻwׯzzz5kYLMM###I[ɓ'X2Q^=v7nPWW0aBJΪTHIIAeפEbѢESxsFOM-?dF{E>V=y3c2y/^6(;} W^m۶0{S@؛ Qyk-ᶐ:р@5(^A(OWWd$ooofk ==xXLVVVݺu;.H. kkk#C GGGXv)A=ڴi3`\$ٞ=ț7oIJ->H ٲgps^1l"*2*UHQ}c_%ń@=kD|RTWf/:&ML6M{,=!ڵk P?t411QW'ޡJwQUU|թS3?ܮ];%FgϞٳ'i+_ ^>y$iٳg!hڴiFЮ?=m۶U"Z^[)KqPl<#SIvfL9йb$ׯ_!/^(w .}b8777d`F{. $xk׮)++v=|o Ĭ :997/"ߺu KΝ#mTG9w\EEEDrVV}N:VVVŋH+o-/"3>$;˛_ߘ.ECcE;j~fZR$i3$CݣZ.v\"xuKMzy}"^^x1s+ܩS+WHؠ˘7o#aؿ) x. ҥK(r))ňFC@111AEOy %aF"ׯ;Μ9S!lU䴼Ts22 >LZ.vLjժrw;v\,"+++ {nf&Çf,۷޺;w *wΦ qkƎkҤJ`mӦM`8.`IRӽH[!QٳgZZZJJJmmm0 ޽ )vsQWWG??~hnn.z6l|[)rP-/`}2 FZ 3gZoVԸ8lWNy>lذuN<7qq}9.BlڴiHIϓ'O=Ț5k܄͛/ϟ?~Z&|2<X\ h 6P]E{, ҥ 9C֭޻woSfWFF -Tjx ,w]vu޽2AL8, H|| :_|Y]`LƍK.sQ|>}Dʳz#=&L@ڊo̞={+n֭mڴ<+x9y򤡡?zw~]zu#oocǎIdi͛7'bpY3f0550˓'OJ$h^D6缊xhZ.F(ց3Rm;w~MpwIgggネp:HX]\\O@)I`7UUU$XgϞժUiӦvI!`'#*mf X w6Gۺuk888^FF>ٷW6@֮];fڨlwwwt]z_Bʉ(AIhyfdbVكj@ .4C9hq\.Yb\Z5跲2V|dggWN:hQHx}:w|M;uyf K !766.Yf]{GY\igEEE;tPeR ߫I&b2Ahh!0G|P]1 i:BdfYrrBo3j(dC2DoƥF$dY ,yz:iqib.H)!44}|}}6S֮]R olf݃ #>Po޼1^|Ç#3fl7s#HOOGԨQ#ю(gv;w8i6nxuvss[~1c=zTF "رcʔ)) ypyժU6ly!D[WFռyٳg+**"iӦ!ފgf5ɓ'LFjw׮]'Ξ=+iW\ɬ(fD*OfZZ.Y -ݻw!xP>8+V$&&^|! Rp2 Hӧb/8/\~ŁSj>|Iy!744|||驙,Gl2"kjjr?֭Ceu~ϟ?˽[T&f;wi=E0\xQKKyraÆ=z@HZ??aUJdff𑟟Y4OdEWTBnnj}Ickk n۶-^4#FԮ] GbP̳3nٲi:РAɓ'cO^lD+B yvD]F !G`.yqւ'Hk!AVFWbA)^o7n`Vaq{<][ e˖ժU#@6UUx̽/_52ݻ6֭[Ǐ/6|={,铷7TᲯ/cǢJ9@ٳg?L~~[1j C![6mč7 5k.a-O Q\/qߠZ.gǏqyw~Rzu8 6u 1^mll!8bRAS# gTr1>3LN7۩P@ѣGS j۶mEGV h9^޽͛'#2zyyAőխ[wƌ<]A|qΜ9ԩSL;t~ &24>7qjbbb&Zɬ^V_} ;%D}rNdwk8 8w\^x7zɓ'+p^˗// MYtiɞ(QFϟI aÆ5k ߞ6h | %c_~#gA.И)S422RKKKBPY-B׮]{O(55!6tВp%7o.`m\ӄ  šaGΝhpA[[[>ڵkHd<'wPQf=^1^)_16@8z-hڷoomm QĮG577GFTd^F6TժUҥKgW^E766`ojJ! χw ӦMC&c޼yࠠP4*F?ppbe2wzϨTVVVFR—:yΝ:QcѦj#lҤɯy]R̼/2π._HU>ܓ~s(}"- Д<ь   o`"@ 255ApyxNbq"8pիoܸ!ݺu!***޽;JYd Z~-<_:u:}4ETM6rOAEJ'F߿߱cǟ~ "qF///#3*ٳgE"722fvO\L YFCd.5KDWy3fѢEb1[>XnQbb".Z޽H[$G7+8)0(I.Z2&9kPlG/{%AÇ`8 ȶ:d./)ZˠAS7n|ASC˗WP_"Сo IDATH@@.))͛7xWVXѨQ#QGź նm[YGX3cx)L ) .PߨWCA VGɩ^:*\6mڸaO {D3BYi333J?~$mz [ne.] s>}$VZ!]iUrxBK***x^r%aQ/^hk...w$=z@YĻ *c<Zh釰0/g{/W;DH|7zA%%%8#ĉ;v쨨Ȩ8sU7,,aÆHE:46###ӧO #G.]Elxxx,X@% y!2cǎ!+W `*fIϝ;[u>} x(++̜9f5g[[:u@b!d&QUUe}NkpHUD ԪUE+xۛr&] F߸qcdA;wFٳ-ZD}#dd ܖ2eJbbbFFĉq.(~!"=p@[䔰)ff v.xEGF}rUfu֡21Ő-?u|tuuDgC$ӨPq\f8R޽{!.kZ\tА9 xTW==M6 a0hJry!噑>!~EIOwݯ_?f5"YdHv%JǏ<[lJ``jǎ/_͛H6lE,t!mnٲ%NyddPiӦ1]]]˃jۼq5ۂ(믿0p2\8wdgu֯_ߩS'uuuDZ.\(&ĵ|L`Sr[[[C͠r">(7>}z7nȜ;J@q EJ&@qIKKK&)k96ltLA*$̥@8h 4'6ɓQ'Or]tttHSrb"t]htΓJĦxkbbLPe:fATr$"-8Y"r[y-9#«^zV6l{^$A8S5558m d21JJ={%r :H<[#5/uԟk*hokwКs~jZ7sژ1c7o$שS'9mgޡC3˥8;; S)-Gb $⺺+WӃT_pϟ?#СVI9Axo֬Yӯ_?qqE\ǎcCVaZGT#FXD3L׳ŋ++˗/%q%ݠ8-[ō9RSSsղԇ@ӦMMs7F}?>Rďնm[QGXoB-뤉ӧOߏd^Yś%glRR'$~,BW$ ...JJJ&M0XF~H:EyO q=QC=z$RyQQQ*WE˗7nٳ'4dcW o߾Bn T!h#//"dmmH8t3 Br=^/^ G@g ڵ qU-v-D7;v`&AtYjUYZ^/Qـg6AI|wS<N!&&5~ر-[rCMkժܱc Kx xg\쉯={ʍ-Ď"/;&I-qF.]y 0aׁ?@AsQDfp5u AJ:!H0,x5jժU~s!Muɓbvvv̀4v,_t  $k׮̽\4df`o*\D/_Asw]!!NNNF>-GqW׮]_]C -IO> Osgn^Bn V@FFFGWW.b<~6mZPPP8^z1N;3P01/t>$h[ll$6lؼy3RI__ߔe/^aÆݻw?|pՙ2uZrJ&$$ԫW޽<`kϟp# GK?~$ɤ߿̌Y( "=z; ߼y.iԨQFFF@5jxL|kĈSdݢjs+)))(( 8ԩSӧOCtpp믿xa;3v񑻻0U( /_666f⸀qqq|8N| xϒ455-===.>T޽/]tr9HLLD#] 9QE3@O -544" ү吢[n!C]dfAG>FzY2DJ}1d $X _.3Ϋq5%- 577Ӈ,[ 9fdBHz޻wٮ];8QEAwBEԓp\̭[b ͛7,DR{{{q#"AAjh A2 _F~)++ 0Gsyx??? h<^ lݺ+?m۶Ւ\#(F^^b5^޾}H?ȑ#nhy!Q}6m~^P{A5C]p:t{. Q 4g F~ܹg"0`KKKҶ7nxQPϟi̙pf~@BiӦ8YE}Bx+ Z"N@Ti3˯n:!V[[ &\_J5"h/h1hMVVVwG999hrFpUĉFŕD.UA1:EDD-3778dŘ}̙P !\___|,+W찗.]̂mۊD/z͚&)*sd՞ x[KԂYd(OÙb>sf Q4iӦ-[B-؟ppRpjBSnח؎q(p:PJ1Q%'p;hϞ=DctrrB^"݁~#NuttDΊ ҥ 0t h>}R" *1@D1΢ ]Fy e1Aʕ+ޝ!.N :thԨQ]kH;<.ݻ"rgŬ4AxR3b4J.e|?oq6Hfb!!!8 iFɌT ֨ѡp޽{߾}Bh d͛?Sk*(;))iŇF(\j׮ Gn:\7VlCBd9!ԝ;w,YbiiieenɻuB~tȐ!~ZnҤ ZjH׮]qLGt~9ySNf!AʸsNbRVV ;v,*ĉwڅƎ$9 w_Gx뙒ZOO?%+rd"<'Xr6LXǛ7o EFGG&&&b a 'D`?B!`]tA '|7** 72$'!!ٳgΝx8i#.HQ 6mED3:q_z1@|ݻwoȑ# 3gw?޳f{E.p|*z!E\.RPڂpĐPFi.AEVpKEEိʂP(V0-׮D ~ƦtZh!?G}_.m$C ~lUϒx&ɎN/IDATn_|ӸYبZUCu ,DY骫c.]v[˄zN'իWeС[lJjnZeoɄTu}@ZgUݺuKK&,ܼyS~.IhQli9rp_K.#=~ݸqC6Ǐ|Ra7LUǟԛmꤶu~]DZ&hTzi烛+)Wq,bߪs/}+Ino'!YH<9riVTǟn:҃lFvKi;"-#~i͕vM5Iǯ\“4'm&=ZϯemHm/Znx:_eNʯ}ӛ:*?ߣĦ-L~exEpŤPs*NF+]~\@^S?fҸFFF Yqy-**T#IGd%7;O>C6lMq۶m>l=5jԨkunݪy.!Y՚Å Zh1{SNI9.11_~9#Ӳ:o/]tڴi'Npy/^ի4Æ ۾}F4^S&Zy+4ZQbkzjVy̘1}b$''O2o߾jؖ_p jih$I"d7~]B?Y͵kJRС(}KXauy CYKg)a?˗/_~;w=7,,' ޶2ZazV^r܉ \;Թ)>/9 =VY0;=O{y={OgϞOIO<!hݺu6mL&8ݻwgvdn~g/3uT :۷/333&&Fr}l6߽{W2`ټlr{׌Zzw؁J֭[οte/W֭sr$ 0ॗ^)wٳgK)777 @ r t l(_~AĒ%K/MmWo&Cd *e$Ӻ;wT*PiNUVtSرC2fyT%C6mڤ>>Ν;'_ҳY)U/.55ˬE׮]Ν;sX_I6lP' ErrC1wzh=jZ=zyfyAӦMkwG'$$8SRcI' vB?\yq95jZx;D=^y Wo[)V б={ϟк,u)n%LZavbyx& @9+r 6|'Or@'{C|3gjX tḸ8#+Ѻ,QLӺl<=/8/ի.O8/""E }@,Hn?ߟOӺ@݈iذaZnzZ@#I{g>uAJ׭[u)J#GԺG,xnngϞsպG,xn0b9޽{˖--<'+fŊi<4'0lֺ ļJ. `x24y9'Č7mڴѺ,ļ :СCZ\rb4;-b9gq.J|||TT֥0k]%@HHH YpP pА'^FQTTu)E^^aÆyyyk]RM[ֺ G2eJ}>}vc6uiӦIM]V날n/X1tR K<1j]\cy|9F,Lhhhaakzj.ɓt2Eǎ\X駟v… ]{vgϞٳkD>Ev/^֭k)w).sŊFyAN8e~sq2q.SDGG7Hjb9F,؈'ޏ~-p?c#`lrX06by%D@eee*32}^RUԔXJU 5sڔ]6D=kZi@3YZLor 1,,$'nCvEPSzhO@[^13U)cͳB)iYRmyiZ ҫc[fmUMLqFF&l9q09>m}қK ~`VrRmEMFmnیPLoJJlɫh3loJ):Aj$mԏ9}88^/W}豊'=3Fz?+Ͷf<*SyoKSʩ)1~?# tYa}Ya2o _S@Z=T֜]+#+ʼtjJWDEFĴLϙa?qbPo\~2=*>?'"~h)^fU5KWKےhV/G}BMFV%Pq(>,jE]ۆ,OMI;_;6))$Ԫ)x읞fFv*3]*ڒ(SD.KrFRe:cIUHlbJjJJrrH^UPNMM7神/(ǔq}0) W.r>uTL+d2]$O?W53;+]bqR~EM\b\˿z۪L 7())VX,xF)#`l}؈c#`lrXqMTGQӐ.\g C2 j짔_x @?[p}0}meKo&/cOݽ{6/hǯǞR}@#ĈW6&[vU'ez}3ܶ2uz; o{boRoWnOŕ{ֿfY 7,&}3Io@HxxxǐVˎ*%gǴ (3|VVZ`Iή^B#;xMܕS@άh)vU>ʲc"ӣW%eDs)o,%Kj:Ϭ(D&*?9ԛ&ا+Ǯ:_\~<N/H;}T OioK!o ZͿAa>#[6 f K5$xh䳶ieJ-ӱ)kfEN&/yLqb끾[+2ny㶩=LyDiXY˛{"pB)nE`ct% g4v0       H___PPT90 IDATx@{((( .ZW[iTbZXm*=plQ *( 2 J@/ %$pIϯ?{w{wo0|>@@ {x4=A?@hQ0 MP1@t4E=} {x4毩eptttL,uqLSEԾ/M=cngggmmmё|14g|ROӉE;̉A^e}@&ш}N@t{̺ 4*Mf҄&}gVζ!s'YlllT^z1iޖl7vu{[gy= K+uv<H<P#5ho'I'Ts.,SP9 y\!=/.1}uΝ;b4]PX\Gtw4!m'{/II.]Kxh=ի+VH& 4*'NՕqN\\\V$@h.]#ٳ媧<f =T]t);; _|٧Oʗqȑ#LX PDwEiiiUWW8K2%-- fiiIo8˞ܹszoԛ7o6<~^xQYYIo:::yyy 7o^_̆P{p~dzddׯ뻪k׮/yܲ/--}s?6mhPW tp8RZcǎ]իW/Ǝ;zAF׮]qCݻwUҥK-,,i {lmmM&RUGUF87eӳgO۷obrEeaPhYYY;Fv}GMMMEk֬<;焆ʭnݺuX 766Drg͚%*OOO/B ރ͛7~Addfͫ'M$yŋ *Y5 >>>ڵ矴W)_CxZ4q۴iCJٳ租~*yr\\aZ?Q\'ѽ:t@17:uJ:88p8bJUUwu-hZbw&tYEEE8qmff&yƍ]]]i_=v|C UVxc׮]NCŦM]psY$w\-ah%y!M4_~%aÆ v]n<O+WP/O)psɓ'^^^ 6777##CcĻ{կ@###&Z,EKMM}!œqS_~/D>>>/ n݊_n$ {Q@Cyrd֎ r˖-{|(4 O>666ʔf =h222pKwI닄bٺtBo-9 @={ȽR-BJ %η,6yyymڴ Q߾}%ܹsM>Șm1dn۶Mr%t]iswSdT@4aİ03Io͛/y2SL/ߋ6p߿?$%%YFPyyF\SN=ƏO ̴MMMUJ$^{ H4aW^qq%Ktܙb %ܹS62xzU &Ƚ*!!=q]Ν;goo/Zb4III_|!E9B-Ohkk+_XXxӧh7ك 4aČ ?$1NTTnd㍊  /Q/|rZz1iӦwG!ӧq]߾}LT#//h Ao111!޹s'p\ghhتU+ɓq#[ƃbq|>/}͡d_3}tOOOC#V:I𗅸r \EEErµ7,1* СC2U@m ރ& 6n܈c80B+V̜9SdpĕСCG!Ƥ{99ntSSS"{qqjć YYYT3Ė.͙3G%VB]#֮]+uԬ%G}7n[nmggG1J522GQz:yM6W/?^֍QL "K#?5iT7N05|pbʛ7o^u%\3eZ4~K?rfff#Seeʼnl69èQH~~~u\?=5u/-sOjl]+MUk+*ݻ utGPDu&'ޝ9smó/_*zU\\Ǐ~w7& 2L{]zbTmR͛7oy~=xPӧkţmCYYmlw ȧj"k5Pnn.Pp\BHrx ubؽ{?Чi{^innQW|zZr i8/^DGlGF'.nYoaQZ}:!Q2KK$nz]֖)>‚sD …K߉[,\D?2nzd ܹr.n]s!?#&=0Z8p S$߼_=D:tёJu떐b$]re .ߵkW23ꓛ{O?'ݱc  ㊅ܫN>}%wqrrZt2)X+K?_ECs. }DJ-lBU }RTwBTNOfdOH &ߏn>Yxq}6N:#O0lذ:Ç>ԩL%7z\72|?)_d @ 7N#MMMҒ{Uxxɓ'?? 4r\:1Wao=eS>,<&ٹ%?0իW666*EKw/=/%ߴmyXXx63k‚1bw}mV+>|q$_]>f{FE+֮}#"h"f0ݻwAaB&[[[kJ\\9UE+맅^/õs 4ȷ~+UZTRRCnJ 5777ZvgϞſosenaCMX"scG͈ٳ!i\`Rbbbb}/;tP_Vw?eoժԮ&M:-'4#_Iu V(/<(naRZYYݺuKU=?N}xxWq"^r|FJQ֕MCvrfJQ;lFt~/w^ͥccƌ!J}ǎyO?ھm۱cp QXDN?PP+ǏR\jp)..3gΟ_{tDA?).g zμJFE1X_LLiߟ*vo"tx{8YWW)Lߓccc}}}%OyYrFGGS-[qyn^KNwRsxb%.!|-nR;۷<ŋ9`b:Zpp@nn:=HGm.Ou [\״i" GTW%PCmT,73sd?#S#x;.!j,)ݩKO,<`q1cWϞ R6oMT==UW-Í|&];wG##XV2Wp%YJw%H|RUuݾ}{Ŋ'kkkϘ1pRѱcLj*w766JIIyrЬ :]{nr~dpuu&Md,h Vǭk>vw_`A=p8zٳǏonn~ԩ;v(wÊ ;;?|vqq 1jLr&2ٳQMC&&&Fu1|pqOrX? q8| 2oq999;pxY\I9rZ&@}.]Jm+L}}֭[={NZ~:uz/1oOBB_W\Mt Qk)ӧO?[O.vВ3nl_aÖm={l~. qm#Gتq2VZS<==q-311dWT)&L[[LJJ'dmmM1Yf-xDfOE\ލD<=3gdLK/u8d IDAT|;zZYkzYg e;IkZg-XD.T9F($$wBƍ7i_T@4a~~~7.NP"tY@c3V׬Yۚ]%9`ٹ\N\"?U;y򤶶x0ޫWLbB*QXXwqM,-1޷"tY@cԫW/fdd&B'駟DkJ>p,QӧO G㉯qcT؈w7l ÇSZJKR_?6J/'>A@ Zb@H.uMڠpoooϟŝE+Q{TTH0ӌĉ/襧G|BezWؕ2==}ƍr߿!R'Lw - ՝A֬YC|k"" =hݻCJӧv3%ڵk2[_`!1S7eH!S.TK]B>4eii)yyy[6A6~iW"Ͳgee%c> ǓWt&I|Dlq&vNu* (֟Dp)j]ￏ@*=ΝK}:NΝ5/^LlyO>]M:o%ȟoGxxb*6lذvZ 2Æ U@bʏ?hnnNRj 񴵵?ct"`#=S~駤'Tڸ]tJZnQ/m4`7}ɒ%. @g TRR742yG>33S4_M>}޽]}}}$JX\P}_Sp>"""**JSE-{Ao1\{HHH@'NMK t{-ׯ_+<_OOO+`0HL|||XXZF&U;w޾};6CcIOOOLL6lX<B*Ovܩ d8v옑ѠA)wuppXfKZ h@3QpGzSN%ޔf2ӻ)r_ Q áJ-#FhHexUT,^8ʼp+_ٳױ]| +B`XX `>}S 4Ub1ZԜ_a-#%&&у֙xСCAAAj W\\NJb庺?={4557&5Xy =fMOOե>E1cFxUb0*++ Hj . @V젼斡k.;~iJ]nݻw8Sn9sd_3###>777/%F-tNNNn$-mx&6û!~ز3ݼy3))iŊ ŅY)7hnnN}:Kçі-[H)6lppph$;rȸqpP^ɹ'RK(9ǧtҥXkaaAF{nڴ]f͚5tPeʧO>%- ^vm*Ν;  YF#5=|wTǂc_Al6[ 3[tС[pڴiG55˗/۷oԤGJ<+&ٮ];绺5LUTY9z OwJ5X4@=@cAz^QQdW=~ztRci 6|AW>І m~A/&MTF>nVWxlc7q&$о@#㫫_>vXWeeeH)mVdf"(/^Ǐ)Wjj>1JDhԾ'Z(zbpQGg \ip{$\_d˫ƿqsW_ͯ,|O?D43ˊi-8S14*Ɏ9!ͦׯ[nMjjWhDp(1  Vꤧ5\H\ pM>#o'%މ/6f{@?~|ڳrN{Zg;cƌ2}htڙmS>XL_u>|@7ojEy>75: !yQ~-X8p%-u|1E˜5X~o/X@%IPJn N%0S'2eʔ}*[( f@3o>D{>--cǎzzz]??/oho CZJpb.h,4zt)СCYZ+T3o%0nF| -lbf~2h֠9Ӎi~fRE8+qرcZsȐ!JHE߿_߯_>9)/R۶muuuŻo޼xǎx riR}񨮮rʄ Ͷd7:tX~=qlz~={W\ZZZ/޳gKpˆ{m*9QګbZaٛ|nS|lp?9r$q`M6^.*(iip/q},ݾ/j2(t۸,Cf/11Ҳ1舛ѭZRŬJXzҥK͍4-jo#xs`;s$q^F7mڔFLqvvԩ#ۻv횙Y]];;/̉<{1B57`_+&9 >@ղJJJ4]Z< vǼJ!faaa^;v͛7qqqLÇ'W:*Ϸof5p7p?O!\Yv ׭Gbr6NQ~nVVn+/jPUBy ?(4~O|>W8jHgΜoJOؿTH%8ϯZ.N XjmX2A(a+\)~(4G sKKkKsga@+ׂTȨQ)>|91۷ޣe&O,577p8eISр &J3b}`X:c} g#+Ġ-)({7Js&ш}N@t{̺r,YD<Lu.]Pvvv:tnܸ֭[?ry+.0iiWeNx{Ȕ;[p2{KBGo'=#0sG-Xl?陨#@~%|*̙37|KkM 4G{nvxȱ$aݲr}XaFY>•^FOÂ~xX=: ,=f1iEWߡ]v)T*-**Vk``"ݽ{7..nРAaÆ4I*|c1=|g(yP̜A.vG1{x{,<%I'SE޽Cz]QQC Trsf,رc*Z74t{fA-#dm7`,|Wo,cMxNyp&ݵesFwٌ9*e A6˖->|81ԩSVVVٚ~С[[[v}߼ѿDjaaŋ)((狋;x`C6ڝbe/&۩'B)ǫۮ;NK+uvx& W+4^ 갱$f=hHH8F@~gii)_ٹ Ԇx+u|鑂_A}{ vg^IJ>yUhf5t ݌܄A.1.Ū2l0d;;;b [ࢽ g~_ɓ2%rC`` }*CxNR(rx{@١ yJd=Ff7Wp8t4CRʬY.\`[xqBBx755499Yշ~ob pϷS@6Bfi6(JBmV^1/+gt%5v|XSytEJP{8SڥKb [͎=A,OxxǏ{}*xoܩ?n_r"!vG(&T.BoDω*:׆m[XЬ7-QF*nZxٹ_~JG\}G333AtE;!uN q0`0LXp2]K@-.3g/^8qgϔÇ۷ot@Uho<{?@w9;>|l=ŒW,NE*1ԅ|NN;8|Xx߿]֭۷oS9Ǘ.]Rtc ~J& Ɏi;q2;$)nZסה;vh077711Q3g13f~VV . 59955֮yUU$#CC e'!I#}]]͛7JZ͟?/W_S`05<|XZv츫6)IR_An:*gK-^dP:2kՑHݻŋ*s|w#셜**,AkT^ϻvڃJj׮*&uپ_1'aࡎ9pLšc]󘧴۶z7ss󴴴Y#ܹsц АznUUUgΜYj3 oܸ1|b͛qr_IHGba ]@nݺŋ;88Sp[HSEzI~p~q8 j*[.\HJ=zCCCܾ};77w ) amҌ只4rjEUHZ>2P:-⬬,Mų޽{ǂ//73hq/D?Pnx;٧;={J\.m۶ M{@c xf`NaaaǏرhwذa,9|ϞK([^^O٧w911Q=H?~|ʔ);ނ1]leu[7iӔMiZbX3gT@njsuM-}*MOÇ/_{{Y]] {P?C {/ݻ$1 *pŋǏ>}zee%1q޽k׮U<>@LLLpō]MgϞĩ^xk!_V"2\Rr;w~٧xrnjCJرcLL iP1̨O;z-k+[XJbO9=LlYsBr%Y|AH8eLMM۷=DR:ttt|5Kg޼yӡCh"6tWrS'!Ν`/t>bB|? vqq[)uOwuKީS% 0`ƍ[OAhAsU 0# IDATZp#.SXJ (~>}:##CKh-ZFBp>77"_pѣnULy@hLs,{7M8X4=Z'S[YYSSS%QVkwCCeۺuk777i IO>}b2ݸ@vtprlm71ƒ?0}9𰳳^o߾666 AHܫ_--h``DwS0@;@2Ago 6c ٳ&سD&mgK(͛#Gb1hV) qnyyg$166OI|G;Ut7f/_>ɣWŨtshU/o߾჏2yƷiiOvdC8꺺ݻwl<<%%%rRGnݫuo;*ߩ< bˆՎ=cǎ6uL+Z[[vG "zF߾}qcǎ@C35]4gy49LJǟ>}ZSs :m۶!Cn455U47{{{8qBrGAj;88|;c|Z3j3o`I`-iIQibDI -+W::*ŋR9 ;v %$&& uI"LrGbp`ߙsc<qxf7n#-~261嬖&ۿu֑ϝ;GwK>@Q4?rp0Vܞ[V-PtL .rt=hvِ˵,ɮĐ/]dooOLf4+2?+8P&XgenfoK666vZW47 &9in#BO}0cEEӉK8;;+Z A()5D*G~ѕ]~r;wn_Mܹs99Ll#%r+,,>h8롬+ƇH=t0e u\BX~~q232B'ND.6MqU gff.5@!d;nk权TGdRf=IwڵǏSL&"EnĽ۶m{!ׯ( xZ大 U>%>yiiiJ@[OvYTzWK)9a#5kܲ_drMLL䞆ke,qq`sKr*-BݻvZ%CW/ ދ9 X^ M@ {{{q777WrG\]>sέMz@|$Zj.](@W7{ {` g0NZ5p@ɓ^W_EԤ>eeeOmȑs&cqG.rc%C[8V3oɛ?qlJ9>^G޽+(S/RdfֈwtMCM>]hE{gZ/ &a?fx,SzuUrl@ ~&ш}N~{̺}W}ѣD:333| }}}-ZT'u%%L8k׮ݺuS0@>R sRHHͻE(S6gN`EДJ@B˂u#Rj2F=#=U|dh^~D)SI^{ڋڼ p4 y'G׷nזcr8exOu^n&y'!L ,\{BcҊ}_R]R/\|9nR+wyV,4ZQQIuuEIG C{y{iO P`k,3܎9!  TJL$- nLkkkȭeeeYYYTܹ᡽֭]]mZ٪}$;hAs7nӝP/h_27ȦYI6-Zs=e9ш'c6M_I~!VgRg}@P>CߡCE/wpprfnnnxxxFF&&QQQ]GI)(@ÆuĉgOE}o5@Kp ;o=s Fxp˦Y'5:׿>*7`Wwo ;4gx;s;i;#St#g&j*b9w5*o߾} m߾=iFe֭[+q9˽'''+**Nŋ_ĽQd${?7,q]%H/e/ͮMc:FUBJu3Q12 q77.]h &EmKUWWki޽{͵SX[9,|4z r܈|ؔʆ)+XXG+/|s'='xMZoߺ=8qyQQQmҒ%5= ;woprj߽zv£ܜB'Wlv\Н2X%Lp?b/:Z #Zf033۶mڶm_믤tpMzzO:{ޢRKW J]x13].<G✣ˬd]J!snQN!s[:]Yye}c h[]-bb\`~08=4)j#S {*332wJ{Mp'|^^;u$NLMM-((>| ptϧX ;sn;F wՀ uʾ}8733:uŋI\.3,--iGsrp0Vܞ[V-Ix+@e寭=i$Imڴ#߰a´Lt+=ϓ\+LYЕZd2sssIK.}yii=zK"A()5D*v+pÿswkм=?NJ7006li*_ggg*k̙35]~֕uvC:>Ç?JJJ JJquĉ::̝;_zzz.E_O0P~k权TGdRf=;k}}۷oLJ7oޓ'Oݺuk36`OJ| tbx{yy4XhptqqrٳgIl6/1*4DDDԥR޽{g},O.aÆ t`0r&ɼtҲeH[n567߼yC[zK|^%k;Ճ)YYY/~:)}ȑ/^444wM&m e=ؔ㡂nxO~h"===I[[f… X:H ۥ)|" {7m{gw!nxTTTҹ\GzMLܷo_eeKWG# H ?ܝPҷ5 T&GWWwժU̙3?|@Lܻw/n⫱t|~H!M{c^,g"DZ6;5$䰵:ro~*UTTJ7Yx  lr7k3&e&%bsvI$w7W{=g=&L| 0Mo:F- '{hTƌsU"۶mKMM(8@I p7Y,$C,dt$.l\ȑ#k׮q\}}}bepzKWKڟxY_%glоԩSWnə;w ɓ/\[7ߗd|t&U/Ja?|;Mi\OvڕؿǫPdUUU׮]ȭUr`Ȼ3u,Ijh\(l)S1B[[hdddf^"$ӦMȭUxoc6rWubv}fΚ.Z0ͷ|?p8ϟ߷oF+5j. ?˺ I)l߃ayu{덓'OSfΜ@KQQQA䂂\ ܹK%Օ+Wƍ'>t{<@ wk>؋#tYEݻwd+$CѣGwءr$''K=Ћ=7~ТA!1pݺuYۻ&|oW-Pl7VUYnY5)\ ڠx[|)t薫[t+Bb$z-$ͤv[&3gΜ33əU1*Q'N:z뭐_%K4oO?=gMaaaɉiP*3A3IMJ@  ==?'6L]?,LӒ@,ChV,UpAv$R7w܁Y0 z'^uUsںu7 6mjk?G~22ȏ?rÙ!/> &ĪP9~fy.ŝ3,[4dBĔ|_ p`x͛ydk5зnݺ|r¼7)))///Ft߇ 0d8q' Q*aJ?>3uTvGMKK@s}"Gy:g~7a'+`̙4=j߿a'utt^cVcP~CEϟf32HKLN';P\\|Ǫl1㽗Z_^X`< 1^{: >sȑ#CnJ?M|SL>O?4;駟ƶf{3338ycZwܖξPchĈeeec׭#ggΐh؁C9r$Edz-ʊAm}ٕ֔SI4xI1,lSMxbwo8Hn&))c1 @kk+վ7@rd.A/;+w$dtzxs0f̘7l >$x?^Xy?'u۾nzL,p~ᇰ_x񙤤6mAmqvFF̙_rzR:ybQ|AIǎOzz:;0f̘qƉ_^=JAc28ZnsyIRAA^b_uWOgbفN*qĈqԱ_H(4 -/پ}ҥKEf͞%33Sҽ^ yi^8zzz=vҮ]OO'L/3f^<{Έ#­RşL.e߾ꫯJx<!pP>LTN5kVTy_z%jMBBիNEer‘#G/@~~b^<8ThbOV~׵"\r!pO<v3<#*/dd?$|s]z)yY?x pz$ۊc|Ɂ+ 3g/E&2e,ZDf=.BLx,048p@|v4*䣀~駟`" 2DMMMEE7xǛ6ꧦL,f^*H**h/(>~Kd2|IJ _E{1xm۶ gªnڲx1Ex̙-[_x_P(gr 7]tQTǣN~߈O/pwed֬Y H\uUӋ=@jjj'ꫯ؁d2tlJ&}+))0r<#<7QOO<m999;x)]A/ٛ0aBaam6۱cǢ+8}۵k^1;޽?QRSS+8orv˖-B IDAT+8}';w.;`2勑<3fPpQ؁'x"?C6%K ĉݻG5sL3RgF]819#A1kOLqӧO177Et%bl@#9'|5ǘ)aX ߷~;ajj&''3Fl[---NE#9# &Mz7p|ĉc92`㣏>?~s=#G-2ѣGKKK;;;]P]bZZZljs)__c~XUοx_?7h48-<;1cیGٸqcl B'|fLNNʊmaX1ƿ,*999cf<=`ӣ3=`qԩǏ#N:#̄ΎήU61bDfff\rG1bfu'oLyI9I7~m I99 mDCJJ̙3`Ύ=۟XII]{~5rҰdu Di24ifMk ##_G΃yNZ@HeTm>sS^ݹy#!Z_Jf7䚧T-Oˋ3 lgPUe9Jͬ|qo7?t2x3EHp Aؾ%}L-['g;0}1ժW,)>?ԘYQ~Z$9IdC[<2-U}5evT(-bqvB74\&Ʀ,bYܫ&Y8q"f*9|/$;}劺w7.A/;zwiq*,@={ɓyHoQRnu&))䴃\wϲ[N6=1ݼ>˔t!(==]V#2SiO 4*=y/F9SjQ꿭Z Hww\^TP!'$$t3 |H[PP09b^Ν;㑳T=wSN#g\@|ɓr<9# ))){C>~)((hooK/ŧdFQ|ҽ{OG?ħ;ykkkŧ֣>K/OfJ%>ڵk7l >:Ѣ OUA|͛7\R|W_}7ވӃXvZhO}ħojj3gԗ O-Cħn!N|Ǐ7N|qȑӟ˺x)))O*?U sssm6[TE!{C>{$<_`X^{ {C>{C>{C>{?oWggGg@c@H^5RҨ{* &UvwttD7`}g:;vv9{O^qkBcăAUj3^zuwTȚ1uL{WYajU5|㥹+kiUJl㥺}xO(5i24iԐL tb͡vyUU63uT6BF 'TG)]f`W^dhnT3&4Ǭe^ШQg8tu%}xcmסb?6WɃ`J]頴`@-(T7@7,!LTRB5k˩GU.ݯP}}=9~ dSُ3EHP5(g5?GKt 7۷ a{I&PJ7ua;vWՙ,V-{JF^^U:J!U6j)5ZFP*&TзiMV^j#ٯP}}iV?*W6hd!TMQ2uUeinnwET[Xܾ==z%NYEۧ%}@.5;m./I*(K;uvu:d 1 .XG}ma}v$qlXQh}}@@@p?նoIIIQwYѴW8Ēe' NNv٥;5~ւ҂p)ݭ{> c]!:woW?~x<OޕsZ]m{Dҙݛ%^0hĘYEyoIsv|r"^"$%f5B 7=@i޹ĉ᳂@ *N$ZyTK S_;ța' doQ6ǭja|7JRViU~=ec0/Jչe͠=ʴVD"7RZ\ puC\leS+-b M+19G?O{ۛj0%W".L7LJJ7ՅPK<_,]UMPk/ϤGtww_1Sܠ~wO>o,Uk2Ւۋo}n1~_~!fϙ\wrtcg}?]{ϽUT;n06,{}uqKe2kk/5t;v.}۱)7u528(H}}=Y\M#&@p>ͻLm 2V2;kXlj:µh6dX4D"Lb5%p1K06'tjh+eMfP}CEUf㰙&"4ULZfZf-ʍ8n7UDICoojYVeRKÔ5Ff5;=*W[/۸ NBb "\~)wPej&zׇyEahd^qxG˕v0kƠgUq]&TM` DhVD 8ث㽩\nҺ@HWjm}NXU ZG؈.W3bUYxo77R^oRgƺʊ L5ʊzâ(&cs L**PH%e#6i̖@2&f*SSUY.X<foYFwv#[>m\W+Uu> [Bk f&¬S›\,v*yckVvRuF~, *z! ֬UT*.$F513'FV21A[xl+뭼1*Ss A ރX=V%sP1?C\;(l*ߘ2=$ABa.W5FkAO?)-=.=PO#[:mή  _dUƖǣe[e"O_U2șy?FkM`^5)Gepp+//QžYSJ{9_`/%Xz! I4k"}?;W. f]\:Lwx$+{XFYXAcb3RLƦ*NMW(MTS? Tj5a>}TqyJlsPZh5>ֿi;f_l'E#;^Wb{0 BJ}5UvRl ʵfVg$l[ȔcY:c7(xYrJS/mU^~ᏉsF&gkT \}U6T8MiQz*l ʖQy_:|:nGkfUzo^Lŝ8Ll5L#2#/%BeE5y;1ai{.G5/ӣcq'>J)LL=U 6nP˴٫y DCTI[3qFaRHgJPiHxњ;ʁlIF_0 g42ן+R|=:E p\5cҰ ST=USdK9Cq\λ(KcJ"תx1g_[)w=}%lui\!zv_bjǚf]*7~{Gcمʵ0몔Z&1^U .C+YM5Epwۢ\FOU68>,yQVszEa yE?(6KS|Sڥ([]ʤIDNVHSVEH0mjn o,wmټՍzo#ul4[xR҂cCȑ{~n~9Z.Ifb7,e|9x~`|yzGtxt&/v@q^yk?KRzs5P5 Z=5韔z_bʑM {{3jt\0?mjNFU5W wmӦzow\ xPu7gery>U͜DZ¯HK-j oȟ%*Wի_Ăߏ]k~n=e^Nzՙ߹yTkDr<߹ku)}Jm *x.Dwx#s!򢱙ζ}s۠;@seiz%¤EGV~{G[zx~1 {@_m,m+'uJJԿc'5\6/;p~KS6{4t~ݢҾзMgj55. 'SEWGykk5[et?޺=yO[Y;yu=aKTE]kIU5Ces7f$ݰO-n~ӷ;w?: Wj:+.('\TDq)bESmg}YUͣ[ T=R6Uu+/>J6Bۑ~8?0U{qϠ֮i#JD؛b%ќoIa_?^OEWSk9}.^R2}U|EؙwRo\^V}׎O]tJOB g506ܷ1}R/J賓 Xyy}CG'#*7ʭ{^|xY~nZʍzq9+ ڸ1]{[3OZC4kRX }mwYl?S&?v%K^%fC,%LvwHrbZU8L_\"R4ՌQ3ε ᎆ͑u}~Mk\7g[kKKk;}io៿eS@Cg}XeZ]sS=mȔ/{oP5\fa'?Mo=ɇvv*S먔\s벤Spm,=ʕuuUhNv#wC`BkjTUrE}w}\1F;y}rWQI=Tu*%u٦SrV2ݷ1rC vu.%K6&T 6{#YW,g[ Wp+&*D]cVÑ5;"2x">"+j70\.o*U3WK+w(?7;/kBj|p{| cD6CУB臅꾛*5>&=}3%=OH[ABejkCI[)&Dk 3ްPU{Ҥ+S7]J@a ໫nj9"O22hA!eutGЏ;-u#S]PkYUf}K"ZZ."x#kC\Ơ[˕Z5e>!*yGXZ!j^;Qap=}G=A;?$Fupguh^M t-ߜ{R|^vWmo_VT{ۖ$b}/r=&f/N$<[s2idV܂^.TEoISҧ\Z6/[U19MNzd!=ܝ'=IYsr ½UbN D?&TsW{o~+e/ZQǨR^5~rAsE9;;Gq wN~LB]ΪwgGwlf[m',orѴ<ݺ5m :jmqӘ:ڻ~_ }/ޒK{B/y?3< wNN7?_]&n6Yz_+Ѳkڐha]5VM~A]58e.#RU2tb @@@@@@@@@@@@@@@@@\f-hIENDB`nD3v99&ۙCč0:k ,PNG  IHDRgAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME /:f^ IDATx \M/JReߓ2-LvÐ01 ʖlYٗBH(TZiK;{_~|ν9緜SC Mմ@ @ :G B@@ @ :G B(Uaʔ)i@ M j@ ZHG B@@ @ :G Bƕ7o$ AIw)@ Dw4*^տV]2@ "G B@@ @ ҥ 22 4-@ ŜҤ3ZtxPA @{_RBJ?pQe]MK@ JJJҴR/'~1A#m UC JEԿh{nޝRd5j۷J@ Dq~vl\_`򅴴4eB 8S+5BnBs&#B)s[-0KKK,$$Y ,Uh@ @wA<= +_{oVc~]@HI#|!-h#B(7[P. @ a)U @ J@  R@h8I>k,U Q[XXhZ@ 4 Z@ Di-C A@ :G B@@ @ :1O^cooo J6֭iAŊDEE=zhԩ@h;(@  @ :G B@@ @ :G Bư?j"SRR4( @ AӧO5-@ EC G"""={/hZ@ ballܼysMK(mJAƫDn%crM@Di ?KרAt ~i(֔|pl ߠOkCs4-@ Ř?Pws]ـF ePӢ!QL)_Ǩʕn]L\,|e޼y>ڲe!Q )M+ m0e ;We+!i(Ew@ ڋψqlVW @ -4xnh~7uZQh?,=z]߾}5-S+9X۝88ZQG %)))++I&Q)a^dbzzX 66&%DD" rJ̕sH眭Ǵ1APMgG lN:`AG Q,WRЎEn/ @ Zʚ5k׹sgM "<?+q]L5'@ wʊ89q #B(Rߕ@ X?@ ZR! 111/_diiiׯ_/Ѡoŵoߞ|#4%-yyy˗״AK@hh@ MKLTT֭[5-3dCCQ _~F.hݻ3K~~~JJJݺu5-!>}ȨQ)aд,/tuuGpŋhW,W\Ŋn (nhCBӲ oC[>{߯ϟ=2x`Q(hV]2qApeFiALnnnhh(eS%g233z aٲek׮Z͂ɓ/,X ))IRh52,00P Ac 9n8- @0t<,r^痞>tPaE*ży&88_~i=##kiG ʕӴ 5jTӂG 2,**JRkRϟ$kÇNNN'$AU1ϟ?_x5#xϞ=;t~AӲh#퉽g޽ƍó?MNNqƢʀ@hƿ79q5BH#4X:::͛7SIbbc huV=tuu5-BъsլiZӨ̌ڲlQ2177WW״4L/nݛ͛7E R_1f{p%gz}J%`͚55j \GW^y{{ U @GGHLL,%^'y:aӬXMpqrԴX$ ͟?ge&C |U33G(CjSSӳd]?:X|EPJ8ob>ﯮΝ;&)Q?~l A8*+VRϟ Mp-[_I_2N>ݦMȞ%"K|4\Whtc f c]vܹpB׮]KMMD '--ԩS4a V= @Rhcccק|Ibܹsիkbȑ#K\kї?[> {}ǎUO{Zc6dj5kaN#3X';OV`IΝ>ȨTtƿ7FCBcmݻ=zTH˖-+[mbb^!00oΜ9H%*NNN*ҢKqF֯_gprya˘[t1K_S+CѤ :>}ںukMKvy۷o iADDD0vX2111nnn%9**W$##̙3ӧOW(99ϗ^*h:_Λ|,osG1>"31Ӹg/͍jǏ4h Hm/߿n:*,fee5lP9R\9z-_~ݶm[b2L6J $F+I~wŶzvc>>B`+ % 7/_'ſxyg(jD_,RI70K Va ]w .'JŊWX!} ۷orh ԪU믿9Қ5k)6yxxU +4tzԨoO!q-'/" @ϟ9/Wfĉܾm6pS!눈u XK@G XƆH٢#͂>0 'O<|0EsNApE|W\rRh?m~FdUg_?+)Pٺu+/BѿA1dftakkIq <*d2TFaddӿѣ7R>iv*T}RH?]d 6i$9ɿ}581"d0;1,a5|;.]("H#bnnQPDEE XQ>~/^|?;&&&F9t\\piB{U*x={?[h“F밅[Uߧj*$b̃9ʕ+7$=G퀃v6֖[.Q؂?BǎSyǏ|+'ܹs6 Đ!CDRp/)pR)? -^a<شi BݻM^v3gn۷M;vH :^|-2D1T S}O8P!d6Կ34>}I&322 Sx 9bv„ LJl޼YvQ<p܄ tvv9ݻw X!s>fhRR"6dUhNh3._%sn'\["&@0G2@<KښUU fK 3)yiKD`>|IvM6Mv.\l޼9,c077Ԯ0ɚ+[,!wrr"D k*G%nIJ071Z~:@LRKjf>zYujh.""U{1?7n?'DMQ2tŋwjV QQQfwy)~3Xu Hmsב]/U "Pt҅(,`KVԩSNNN$:΋/ ~\rO6lXd[LBJU$CW:oߒ%۷pСC#GԴ|[ٳ1o-ej&4nfVVxFʞ2_Gȑ6SZ-ď~^zc1jڵtШRUs5hP!<׮]-[\~=ȄL#$EC ˗w%^G=q t)L###IU) XğFrqqQ/\aU@bmm-R[C ɖd$Y$=_r{Ҵ[i,Bziĵ_~[ ~7|MVqiJF$i/V$P%cbbaXy;|GDGxي+Djh޽d qp9|H/ҿ/IiR 5k'A }f3mOOO_lYQ^^D۷o\ԃxW@>?zjjNNNxJ@a6mDŋw&,WL '`*7-*T>^R|Mcǎմ̙3Y'OiӦ}?}vJJdMZۛI7o<|m& gV^]V"hzz:$X4-(L2eΝ SǼ_ $#::G)S.<\l^^H3iu4?)y@u(;8q~ZtPn4lSn]MK:::6qYy቉ȀJ*5jԈ Ai1!QzuʝWӧO5ХK5kPN%ghݶxbԯ_aaz࢞}zϞ="ؾ}H.\@&/)W߸qzhժոqhffVF sX0`ISE xK,qpnݺ)СCǏs˖-#IiԿa;n2)^t2/ ]+Tf^F 4_C(rۜ 5+ ~7cccX_Nի  NXG?$Μ9j`' 2ex߽{޽{3`m@osz걭꧟~"sPJSB5oӦ mժUp7:r@~ f@~~/7TZ5]]]x+)P~m4H4t|… x:ƬL2.Wgܫx8l6oP5var9ɍJNNU~aܛ)?W;jc"--jժ<c2C)?CŊ} mĤA}恤$;::+/Xb`iiIر#Iн}R ##Cp033XXX _u^pۉY6+=I}]ͳw1F$ :.ACK|ɥ _i)hϞ=_Bdm֬ٺuT8 g O̊k͛+VÇÇg"c+ 4iD~PPb oSff[3Xnack#n Vvg&%&&ʎ̲WǨʕn]L!K/f "gEڜWVg͚ES 44gdd;wA%P4E%v98gŶo.ҿǏCͤd͉1^ 2&Fr޽)'wN>ٰaC ˣGϤIY`QJAoܸQנAEǏπ]+R[ܸvXW1Q'\쀕'|4<"vqj(75)qӠMn")) 4"MP't3M؄\fBOL@'99"wrrb[ TkԨIvv6Oܹs9Ooo.]0$%%"S%$$0׮]+xs) ~ǿKGܿP q -- ̖" /;Ò`m\x&*qOEGyYVEhZ Na…i By後-D_ԩS$8M4!oCǏ_t,]߱9#Bhϟڢ1տ&ncRnܹ%]'`0So K{\7 NBD\rF0M %xj]AJT ?#>CѨ4Ğ3A ٙ6mʶ}-[2۷S>!G)%?`YYY1/Od=ٳ'Y߃%FآѣG~@q00fΜ) Q~kL!R +㑈CAr&ބ0`?7o83x]X6OJ={4F~%6/M_ ׯʏ6o,eviвLJ?~Μ9lSr9G٨Q0CMLLm۶W !CIj3gp 0gӦM 4cǎyA>CYΙzjZ"meWﱘuD)'NNN){ڵQFqtj ?uAbccΝK3 7ykgÆ x\pVK̢EÁdRf͚LZʗ/_~}"4 xܔTʀ <qB 0]v=z,T,_رc[lIvɒ%kמ4i!R̕C=(!)1c .]@Æ ]vM:U|ٲe[ʪVmDpL&yIX! 3 jՊI^ ӧO7I1oߎO/,jVtlѢX6d ȒBRR5j B)moo!aqF9A3{3+#U4aqnذQ^ 8x -|X8pH+ɔCU^}Сy&^zReG`HЂ7 .ƲڻLÇa1m&s1ʊիWAe2p֬Y0?P}ĉP kZe^L< tuuJ$ׯ_ӇۤI %&&j*lpqsb*^1[æޞoPUv@(gϞD+ִjՊ~k;vXPP<wߞg``|.]Jݻw ݻ+ g }}}IOr5j{o)g@{AO(p~$ 6l߾=[*E'>}8x ;nEBV)((HKKUfM 766S9P666/^$ |@ "{ln¤<=/,M3LHC0|3H>0o8}՛7o$*ƲJN򑑑6m  gʕ`թȑ#D!qb>p[ gu !RSSʃ -bMQЂw‛jwΝ;d߮];<`uK¡}O)$[w0l1TٳY R*xnrL1m۶^ i/^gCZ%\PdFg޵kײ^ ?~Ǽ#G^.OXiv,9' DjW\f1ZVڭYk;gs3]Kg I(C !A:~[ 5"HH5(SL!* $,#?pQ]\Pcs̷TS}(`ĄQx߿O&AC4;Ɲ}% [|۷S&poU^ O9ZtW^/[9lժ͛7W,v֍<ߥKaϞH+򰓰lܸw`O˔)#Ƞ4i҄$0B r k~t;ki|i/ i-00*sΔH_(l h"j^wGEr" ohh(;G@ Qҕ41L )|893gغu+0< )ŋ_&D?lBs,_|\z葀Rѳyfe (iكoݺP033sMشi}`@=g( U ˗/&aW XBDpȑ*`?~\2dP=A!OL6wҬ> yJ+ 0%C|t׮]1(~ tuuasTy2h ޽+V0gΜ3l%00pѢEǎc-@LJ{xx5w>r&M,Z|t/!DsXpBJ c|C^TW TԔ>{NDjժcƌВ%Kʔ)3{lrO>͇"çHfa>g"mZIzl<+3F ٳg_fkk[Z5{{{|8T|>NůMKK{PXZZO9A>~J8z >f?قU],nݺF5j`H,'ŀǹ"Gڵkxm``R.ygrrJҧOnnݻ9|LJ:B bݻ~kz*<fff)Hΐ` ZR-73wL'gƄ^ٕ۲M6{Gp lpp1 )EXy;gԹ-[dէHtuu9dSy`nn.HjܹGV;&&F*6lr:J 0YE>CorɈh۶-ŕ,FT/TrʱxA(<0y$1KJV;1sLxRSS}ùP6[>| R`%ԿI{QsGs-;5 B*WLDKйsgLz+V໼$ԩS {`"Hzs&+$fѾ!n:tPYl3xD&QL"f@bڥKrΧ3*VP`(2G)J#@#{ʕ`pi`PN""i̗/_QtQ .L 8`Rѡ֥KǍI)P`yUE R7qݻV6e|h` xޞ%Hۑ!ÝQ999"VOqG]iTjP\i&P5!f1|r*U =y#GWZE) ":PދIAx%'Lмysb<a-llcH i7~׮]v$yqu] q333Acx– ҥK355500`qA%MiT*"ϧOZ t mۖyQ&SdS2xzzqzzz`` S߿w %_PQ޼yj嬹9 !G s,,?s ~ܺukJ)z%ZjAOWgh^1VF) ?a5ͺOc={dn}T[.} }EBeGL `Μ9*L6e2:ˣSN*sሄY.]y& #"> 7_J5Lvll, ` Y_n4N 4 6 y?##C9; MHHLz:v(xĆۮHQu*oeN? * rDp=w x|رciADIDzܹsG|_˯nS42L?%> ]MڵkGx~`yY(I4x`ɤݠ< ǟ8qm۶mp>Wz{{KUg8`r~11knnpH@w SX~Сx#QyfL$7pxIFWha>Fۇg~58ܩ$b;utRa$G{&|y;v\|yh_UVKay}TO?[hѢ}1N:իWW\p >}Lb(_%߮Rf_۽{7h_π O?/ ===Hؔx˖-c'faS*zm !=MOOdeZ={\rb@_Ē!FgP`t쁒^zv܉wb΄yaQQQbBCϞ=9pG6u?>n8/^pL+դIQF{Zj =gmdFT'>>8߿+^^^H ä?eʔw)z ڵkʹBb ]'g bn3o[7n̿rgJz'x q6n1;Ά<ҏLR=Ț̽V25$>5'H x;MȮf1*UAAN&f:|~zbXHDYECt3s~hty/]#4h iZۖlwժU&BE\8?հ@F/Q7?#۷gJۜڵ^+0[ OO15;zA@* 9?Hg !!|fѢE 6eg^@DG_G9|y xZ]ԭ[_E^@7ĉ X]233S[޼yʿի¥ApI-pD?$Rc*-thNNJ}lš@C %}KE89sn:Gun@ًRӅ= })eP-7!~,bȂ pjFɣG1EWYO է&غu۷-ZP-Bff4E4L0gxT:÷Dzyy^鯿JQf CI7n,ۊ+$Ϝ9~ֻwoұcG1Ł'd彑.T޽߯Oϟ) `,8 i$| xaÆ1 b6,=ʕ*9[DKK0wn 0 AAA={d{ ʀ N-iii ecjj*H.\H,QC/,:wG(ײ=zPU}?R'gݴiIH8,(`, 'y4iga6Aan6+ E?O l}=$t.NR.[{.x4lruFuYZ#(OG0&Tw\BW}\`ABr9)8޹sUe|a̅ oΤ$0Pl?~8e9mܹs x˃EjUu]qqi#G&XQ<*r ~0[UbРA˻d5<<|oڵDfDFFFrʗtjO}/+Tjz-G{hZ{s;ѥ:yL^^^]~lS `9MN!A`BQ$?j(3_QiӦ֫y{E =+ RPF ߽yya<nڴZ5y~5˗/C˒ܽ{w)@xxxDEEE[ O5eXf͚=ԏa^^^o d>yEGcĊExLQʃ%HLUO:o߾=k`dd$y(fxΝ]cb1ҲG騛:4*Ƚ߻G X!IX*j ;=m 󳳉=E4ۋvkgr\;tgK}(5#R^v۷Ͽ{̙3`2!CIII?3 g:uxg\\-ׯPy0t(Æ\6lo9kr:q5kOɏP*UlmmUtRAaŋ\fM;i,9uٳgg}#GT$| ޿|/ժ{U$;w駟/3fHOO޽ ܟ ?Oo$pǾ3Vp᭙Y^ݺ_AN ?ѣG6R^D$VM}:sڍr nyؚUgmqtw8EJ}Wtt޼rSl\:yMC1f*;F0j<b!W@Dc@rrrJ~ c"}c2x\aGȅOd? /c-r|2;;L$p#N2Plbg31??8:8|,%ֈm:d4kV gDDۛ7?)V̈́G+cD|忝uv=~1ޥKkegg7,Q sgk M(’Z #uYu`Сuqc>GѴ)eH!ҠAI&R5W ]p-B0 R^r1Yooo۴iC W^bO.sSbÔ雯v2SV9V^| Ν;)ED(PZ<+CzmF>~+^!UHRVüB7B''#m+UXrɓ_$%57χGlݹ3$V..Xޝ;w<;8n,VxʔOĂG#o=x `8}srȋ^ի3Ґ ׯ߷vH}'.o,ɾg)N#/$ |>22ݻğ6l8Itg ظqcŊ)'ɻ}6 2Hƙ3g(:ڹsBUj}1c֭[Ǩ~szݿ㏙=qG j7o6mTb...JE޽->#K F}^HrVZ[14vԯ*y-[Ȯf*Sϯ[a 0n#::߭- Y*'NzzزeUA@@0 bF͛7BI|],v;&'1Ïޤ˔)|NOEGc9K-]nM˗n݊.?wsvvTe叄P{=*Tptuၤtڼ9a2eYxسgѣ{:oe8jn]LX,Ȣyd,m۹]4x$kQ> I*3T9٬YX]+~Y?ǧSNŸnOsP>/RJ4^`e\ww0W\>r"acMڂ.ÃEExBٰa,f Ν۰aCo۶mǎ/\2s?(evr劻;[,Z)m^ 7k,""P.9uvd|ǎ*T(B [ѣG+<~۠͛/_G\q jӦ#ݺ IC*^ԨoOO7,N1-ulRL3+/^d?~USN3`8]tټy3nphTooM+DYӧ-MA GQ ޽{'N[YCRB](Tw+&%8(s%F:w@A']t)3 777Fr-?D7WfLɓz\_g%f*O,޽{QF~:<<ǎcy89/\w}wڥk\lѣ^ڸKD O*Y"ď۴iӧOL~ŊgsHfXh6yrfk)]L٥y}5 rdz>}:GQ ?]vU9iHD_?\\\@ (=իWWiiil+~e'M2rv51h G 4x-*gv0(F*T̗u>JM;2͛77vFwk~5SM du㰚f'i=NR2bOcf0I*੻s:^Xn(uֱK L2h!]Ç+a`ӡٽ[9n͚5f>wSA3F=SLQ=';{zz[ qRJ&&&lvZx_g m6pï?*[`_SA#Z}|"?Wdӌ3K`ݻw?~?NLLezIAAAvvv`ύ7;sp ۶MGdb"'MRE2){MxKR}uvsA'OOXD#GGy6 8Rׯ~g)ݻw9ǘ^ڵc,X#F: `gƧ'C *TuP2wbGJ<94˗/'ưab9\? 8y4֨?|S.] y<۬kUuX &~V03ۨrJijV~eVA3q !尚Fd`Δ`6 `̘1'N ?ѣ[UOCaBK E;o~ ?5{%xoG}"33>ˀ.}̪ڊ+֩ST`t+uҥ]vʅ)N7u1--D~BELzA5L]vh "kr+/hтM4a8 ƒTFe!6xyy) a); h>%s߿_E8 _[c_D͝;wCx'O< -رc̞0S0py?88ѣD bCMrCQƿMR0AM(j1™իiݳ>m/컌>+܁S.M4vء *ux޽/4]qjpSx Bt IDAT:;AI/F.tLQy0V^MuM:a++WnWyXE2|pK.:tPaQBG06%9cbbҨQ#&%kԨ|Aڵ)8YYY*g^@Eǎ4nܘ )ҵhr+>}_>9)7n(*ڿ }^"e.5a_14|1jFבP'%=xAĠZ`)͛5k|A+H$;c 8>ewpQ]_ ^QԳ"^k.j`El 6T Ċذ*@Р .)sszæuߡh7ƗgѢE4!u`<F)ߧR,Kȸy5nq΄WX$.lTT0 /  b$O:`_2"dht *u#jΝw^v-h}P \LXL]vЁ P;x`~H9sU+#G$JT>Ҝ-[\1̚5thIA0Jrr2_'vuCۀ ߇b:&2gHqNj|E$2IiӦ\r"A9Z __ X`ɓ'loA " K227ҥKziӦÇUr>dcam Myw@Y;IdɒAqqqa*U zEG_r%FrGǃ/ˠ"""0gffhB'W_8ɣիW7o,cf;?f٫ۺu+ӧv0Ųcׯo޼^7oިj 3Hs+gU{IVwwp}L.?Ynݝ;wi޼9gs.,XhdܹT`gcE ~]<{m]ӫV-MO{bmQPAdT<}'DP+ZbF/B9n8AWɀ%:q0 #?y]bE]t5 s^@8=}OhxB_xرc̟`=-iҤȑ# kLU!NrF rNHKKӵ`gp (gb8 Km"nܸ]ˋqZl,ٗI>ib0.rNժU+SRZ;cdxccc9YѫP;A@3M>Uk._  >>>Mwxg"7A]נSсd<е{n 4ȑ#8gJъ#@ o߾$NtzitDrr2?6.b!f͋/ *K.h|㘽 JFf!nObx2LpB2e5kFѶ2']\,Q$)JV4L[܅z|Ϟ3f6`+x1akc7ߘ+k֬{İ_`ļz*~]h }uݻn6i1is;<06 zbd(r&ޑ I4:u*$$DAp׎ҥKE fl2xt<bI(%EBxyy#hn4?A^.]3##?pSR_#;^@mBgÀ#Htn&裮(ڵkoܸJЀg̛yej@;1Ք)qѨ"Y_}^GAdmd @+_WQٳm۶E@M>؜|}W^|Dv=5QʲjI~iNj?t,2EJ?{.M*[1qBмysvzCI}~y6III BOOO9jԨ!dUǎ &u9Z6u:<5l29zӧqܴiA믒k$2l}+g;ӇZZ=?\U#GR;win4#cZ9Z-xToV)Z?0(#Am"altzF:vBcw 'v޼yEDyڬ:>~gϞ]h[iueqΗmhРA&0tGhw4{lv-[ :T& ٰa1cpN1"ᠿ=bv6 'N0wر%+#K׶W7ΰaNqprrb"0Շ~(aCDDi^N [f7 7t͚lVm}۶ZQ4hP{)6)(YA;ѵƺuD~dGpcXxG? ˗/"Q"CBBtMO8  1:TF0ڳ9O.P/&;*''5,ZTb rcl`裏s^O?K}cbbd\ *EN4J*ɫa X &(* Wgffg?cyٙoܸ!~Z&p1|wuwrnݜqkj8w RŋS&6U22CexDޚ5kM̚5kו^P:Yb1 ${̴dH*q|}}Νsf-[2ydӧOA\(wp%MjQghF[b٦"+8}J^@2;(QBQtGzyc+ABaߐCx_8ۿ 2M6S^^^ES F_2dr@:,xe08 X~ޮ];GI-]GQ/O.6"՛7P}y$|ɜ yUtgϘ?gϞqFܚڷ/Yfpعr sOk0 zU870D~3pE0Y1}Goo b6Ick{LLLjj n>h3pêUPR-[t֍իW ܫ&)LpNo: &frT.jߵׯksŅdeeqA{)C5]uի(4Μ9B,7ڂ ~>1LO* ,={{>R(iYoٲe{{{{|sWA}4c ^sR H 6$K7oFa{ɲ ݜ@~!O,Y%+7|^Mƅ'2*eavGh@b)Si&CBgLhti~Mp*؋hժQ07o^V-4޸q+WMH()BQxqJX0qDE~X\9 cƌ2 n_afa,7 "\L¸a0dFѡz"~M$A/D @dPtZ5A`<1|P;.R (pD&O>RN:WG^2Am߾$sS(XZ$GH&@FȄPfU,KI Ii \2m'X >FX䷴ 9,sҥCeܱc4u>|#`aܸq=ڵ+d`4]zǟ^#fիӦM3@Z ꎮobBKC*;w5wAL7 ]رc3gΔP)PPN8~~jQx>UILLtwwϣo `6ׯ_3gw YM3cN)_X_H2cvl`‹/ҥKo={I$j @nCM9}tY{lmŋ PeF|GAd-]l Ox0DFGGY0&/+ZM7_ު%mƜxsrrj׮僮? &eUXk~{qo^<> -W\ڵei+W2 L?F7Xb7CMi( 6=6}… 8E͞j*M;RmRfN֭[;_ZBaoҥ @֬YsܹGRh2չ ok׮<~xС2*С}ĈN ]vq2䚔#Gbn?{lC}&'A"~\ѥJ"xYsۅA}Q1-III(m=,慭3Ut@gnH4.Ov\@3,yfj#s;P sJ۝$¯y會h':t(_{G/>m{>rϞ=; d}e2f*uqtt\xǍ_bv͛ o-Y"o_q@0Yu%>3Xs10qVB5?ҥKAOgb={_cW=kRdIqͫW jf$Ę1c ?T&V43hezI&իWg,lrqㆄ`8(zXt|fGTrεP_B@XGqFh׮C K5sڴ[bU_g-10ӿ[qG 6`Z=B>tK`` 7o<]ΞȞJ۴i -A?(QbРA'$A*۳wkP/>3g8p)ӠO'G0r ΝeҥK($`TvBH2rH ?++6&4&mlB~<|DoV,W2)2]Ȯ9]ˉuhFW-[՚ g?A5o~!0'>.P/DE/_`H +Q ÇT͛7]/([Xp ޽{ &MYԋdf4<0iӦdZGj_{QN;wl쪶VT֓`OێQT\T2~ڵΝ;SlfΜ9 @QBCܿO=|HZ8Ɔ*WfnիWo;;hfB(LjzѺ]dlٲf~~~M6Ō=ѣYf:5"e¤C"fC[ha|tpp++W/M SfM͚5\f: Cf[-s".0`{=ejC ¤J\{sÿutÒe[ sPOZdגK׷*t(M̿E+_<EO>>yƘUܘ?мCEGrlL)k_Xz<1(<(7/? *O?!{D*.@)@0 J#K2C)~o}ฉR2e yOK*gŊs aoX>a[{ X˗/g'w_GqPQF7*%x=[$p=fHZ##RMI p(uP ^=è,v_IHw&pG%Zdxbtt4NiUTA`'Ih??zh q_xy_ IDATܹ۷of+ WGl߾lY%~E0H0w\a۶m]tIbu\vmĉr|W&MbEFFw8|[%AAAp .ܬT*oTٳg\.$B׮EPN*W-5j(^Q?cgrF9ذ=@F(xAUF}h f8>f~a? ~-׭I&'B_~㏂7$cMڵkٽ|AA1z4)Sd EBۗN퀻m`2o}׮]H6̘0 #upO?nߞbIzF#s̙` Fogǎ666G6Ҕ8_QJj֬rJji߾=:[,Q]N}r25 4} &΀ TREoc % Ç?ƌ&}%bDDK#@3 :DTTfK+yϙ3yoڴ l {#۷ׯ_9*UQ_zu Mt/o9xoĿi_z9Z)y۷n*1VfϮԦ US0)hFݹs©S2?33{ſaÆӦMCￄ\jGƬ"C(Iv,r<#z݅%^(P8w\]*ڟGe3pB~2gA_DhrROffggg@uST)Lwйٓ6l`H۶m`|txPe1bDFgpwpllf;%/JPǏ+wLtsqq [ nE߀t666ǺN:%>qġCg0}`@PL1"<|>|?}cǎItiڵȢb5OMMe߁zYxfkMKV+n ̟,^ A̚=z4-/_DvT~wAѫQf Z dy- `nܸ񣏴Y^(ؓFkZDD~IS8P7W!Oֻwo] W\9rH"O-[w)Cjjܹ-?~m@"uQ޽;x{ǿHJ(F6*T0Ȁ(qe+>+B2p&RL'diאGdP8s]idi#Gr2-qpp^"F˱?nw٫ĶdDnB\о$_S:?Dnܸ>,XD .\v$ @~3L>9R͛7()+4Їz/IMMgj0 S3C )Yx1NpӔ ?_U ɌH.-s1M alڴiΝ3r7o6QE`AL@>ڵk0ZvYjw/]7,5M/ m&ᅿ%ZD 7>k>%DxbQa!q]nl"@2# R4ݻ׬Y?grvnnn&5jn&L\rJ*Toas# 33&?n BDDٳg'LׯȵYt鯿>U*#N /*fc#9Ă)@ye`׮]]/v*WtK?v[^q{Æ _} ܊Ν;?Q"2qy"v r-aaa ęfE<aÆlÇqq… M$q(YhX?w\e)-&&<ԯ_<2u `ܹs( ?#s=8[\]*U$ԩSLj24mv׮]OcT:Y6mYnゑc*$++ s5,oxz -,;D  D{ԩy+`%AAA],Ғ֭[^_5۷oSF`Tp6ٷlҠRva"''G Ni'OD3x^2fGill,?VrTFSL17ŖG{ac ' 3XrC~:;;0fyՋA.]Dְg߿/xژ1c ؞={p|y/  ߙL۷c,/jMxzz~~~W7`P8g?,J# ڄ<Bˌ~ 4%Ԛ5k/^Xrڷoϟbj*`ұp Ư ?歀Ю3gcnڴ @Ե8::\2?~rC9/xE[ڵɇV~}Nn4xΜ9c z0fgM $F>hRRȜElDJJŋ9ߎ7n޼y̟*gPpyo۶-Zn>RķHc߾}/qԸ  gَk"xxx7EDchѢ0`&}_;{d7揿mj7|جph4bQ8/26l$+0A义 "f|&kԨaJAU{&HN: M%+ v/j֬fNNNħAɢZj%w@ R0!ъpNV?`|EЉuօB T:vx!S`&䚼hݻw׻ñW%&R ,^LjJR)SrWpl '\J _B)%{~„ ;w<ԙ3gߨQ#'I2/_ԛ/pne\ ]}z&M$@A >wxI˘Fẓ#—_~`ʦ|䣏>⯯6fCW۷)zFEl#Gye oZ ]%` 9~-iz ڻu3g0[¦\rK.Y *?q  ooo @=z4 )]8C($|ݭF!晣F:xU[,{ٷo8{Cv}39ȁ͛7(Lz|[Y$2\%~%Ə6lY%EQܹs|9s@NIISO?z?8;IY'?}=:t.۷-[A/0] SкukDvFd!7-<<ܤa.]-L@GoD׮]acE-3)!!!ǎ9͛ݺu3"38:Ǎ/VFܹ|r ?+ ̟L6ȑ#['N8}zp6u_Rri4~ DGGzwޝaÆ*TpssC2N+h6iiiVXVZ2lSڵcH1S8.ZhѩS'ӕoC>LI w>{̠D2  >$Wb6}a;i?%wPPۯ5k/VF`H7sQ(_oRqҿ^m629Apۗ/_.Ծ|o:ei-gɒ%~իW# P7oD5 [P7d`KHHѣUwDc1bĈO?T\TϘ1Eȉ߿O>ڃPG0a??9""6cȑZjpO<{!~Q E TƊ~???# sXHIȷǏ9<|p q;pf-Kc<p3k׮mڴIb.gϞU(/4y;.z⯿ f͚"瀚S#ј=z/^LJ28GfЪU+&7RoÆdeeandM+pCr~W40?Է 0qK-qŊ~kmm͎x]2oǿHGp!Rj~|7o0C۴i#WiRdI Yq  ۷o/"z 3g5ӠA0{=y$~]Ge u‡/޽{` MItEQ,F-Z4iDr7>7odr1ٳE^~ *p˗/ @^&OXuַ\rLt?cQ?LJԩcd/^0a?2ep6l_V-~dYŒƒ #[ll,c5IVLD=< rvI&U)`{9r$^d׵en@9u>[NFSx?ђ=SݳgO3܂z+Ci`%k4wь"jڷos>!!!??<<\ds. 4}}pqsD5k" nG]?6lXR[32JΏSboѠaܹsq1uٳg\`7n5B||^pQ]`G9$=Afɓ'M6?KId[gP3-``Tl!')SF+)m۶1BB{ŊM\o/j _ɨLǏRoEՖw['\]]<^򈈈/hsܹs+P%9۩'NX2W?,୫SN5vtԆC F&V:t8 eǏg{t(Ŝ+VK$GISp<0_Y1F%\I˃y*xq z;u8QFrem~?nݺȭ`S@%M (ca=A5B|SNmQyFn/=~|  v3ڃܹsxx8ɓ'AO~&9aaq[NrEYoܸaC #KwGgf~SN|Ҡrr.22'NGQcΝ7) )^yG _Șr%c9%JD;:rttczjB<44>ݻwZ8 h[neܵk CunٲspرK,aT\b7%TBR#Aަ+NϞ=ȱcцO?OٙrֶYf&m gNާ(&W^]T^?)FUR}A;n})>LIϕ'Of1x;tW_ah-+W!E@Gm7YdBӦMp1M61Qy eϞ=L`BCCoݺX8 !xPA+E%cTTIR(aW2.}O?UvANfZ0j {hdTX-~\orɱy%K"7yG:Qo `AAA#G4=,\]]W a8;  03/^|%ܹЩS'fE\\ras d,Yb GNNYG=W\qㆩk͛ ;wuiʻw^gDibݑ"]|g3j \?b(V?|K@1MdN\3QD :t(ft>,[̘9gj)X ’f)j3f0@3zDgj PP*M.@)T~˗eI f=U73gp+1f/d!, ÊJ&raJ M㖨eS||xub&Mm۶5&+㜃(d`0ڵرcU9)6<@ׯ\}r%·~(RPJ, ڤ\A1w+hذa/'**= &z`ܐƚ5k؞Ν3FɸG+; QƜ9s]ȃuڹ{Ԩ`CgW=FQ ۽Zf7G<<+ DB4oJCBBݻt"rN2eCS:t(`}ӧA$swԩ}̟5k# v{{{57mٲFO>@㱲\m(=lJ(|ɓ'?~&sAEu(Qcǎ~Ad Q1fAZefkkkWHa8EWAvJ.mTZq4?nݺu݂[z1u Wo"o{fܹ:u>spH+/_޼yܭx) ̟ <[n"ܺu8AQH,YR0e&s̩[.a ?Ccƌa}葴zA իWK6K<:VZ۶m3uhժUQse : »&8?ydt㸸8 $9SRRV\i7EOd6vvv#F@kϟ߼y3fV(@P#ĿIg4z\, Ç+_%\iڴiw~ 炇_ŋ#ZZ_*_,ˑf1],bEaKVZ=g]/>>jZ ^yW T 9gx~ǒ[%h׮]˗/g0YhYre 7fcC<*GgϞ7Q^x1tPYzx/e?Dc1`N<>}jpJ?~<{EɭjԨիWYv*aړApĉb^h+++K׷3&G{ҥKV|m۶e) Y2ޙݻA?ST@c.#ލ>|lw}L^%J!veeիڵk5)o޼QRBmH?y?+W,"):6T-A# 7nVe?~xY"H+e˖?UzeEG䤥y^.aQkH@@嗘I&ɛo^xd,S! 9wɓ'-*˗U06!Yn]֭D%NBB‰'FmZEQoٲ92b=;Ud~ooo++C:uL)233/N& I?l^xqq͛7-[V0>Ν;#O> 9bccӫW/s7NBd?A.?_<ȑ#qd?ׯr?>|`ᝀ,޼y[|d)05׷mfVãGx@ʢGx  닧ٿhh(Ni7no,ѱ]vn<޽[oL`E?iٳeˊl?uꔭ[ƍW^=j$XI(}` ^xqڵmۚ! _6g %e?믿_4ٶm~޽{ntΟ?/)}>4vK yuɒ% 9=z 3O5pQs5wCٽ{d!$pHXr$0ngb}gs :D-.MxEGQJs7uVϞ=;4csxt}Wk%[*R a :t@y7 f}|&hSԬwӛ5tR]_}wRRN |l /fYC&{]_:uP@Pԗ/_:88!» Tc.&iSZ8ב#V'qqqO uڹ{Ԩ`CgW\, KȞ^M侦(Ub~7CoooC`jZje&+ԓVE\٬v~}U̍9|pnj׮-s@ wo?jmtoO=pgs\"@ r]XBQ/ .l&@\gOW^›!..rJ  6,^ оQSr)kj"ߤ)}gĉe˖5wCOad@ Ø@ [7D3+L "HQQs ٳ2e@ X>dƬݻQFr7@ !@ EG  @ 9"888))ܭ AǢjժ~[A <?@ 7@ .E1ߑ#Gݺu3cK@0 Eqk׮*@  tɓp:t`ކpԩtVXXB?=V~:vn}=e|QEQP!@ "@ 9'AĿ$&&:88(SΝL]k׎S.o6nܨL]Ŋ) g˖-#GT"%IHHpttTyyzz*SWZ㕩K1Fi&eRڼyQHF'A?@ Eח̮Sz%ei%ݧ𡭜9/glW 餧,j\u*Vk5F&eaؖ,1׊GDžhݠ,SSZ cI<7js3+*}(HgYeJF\ w-'ו #:raJ t%wZQq{gV'?,&gW ڃTYJjVIHj}cJ='.._ݻ*5W3 @Mw_*V,5G41B1H )KU~;YNgβdtQ&ӢţnJr?_([)"6)*>I6X1ڎsѠ.!eidi=r $$h?w?SG+rM vn0szβ$4e1׊s@=C~n* EhB:Ȉ fy*%w)#+ϝ5i,4ⱷɽV g~HKY;8T6β8SgQVv>)|AQfA:˒;z@ E@ @ 9'A?@ E" @(rO Bl#EzܕO P5l\BN2O_L-ܲUna/A[=*֮qiM;5b_jMMzy.\M|AWZ utMb FNܸMjܼo/FV̸5;3=*SpEYV(?\fzzMiRRZnYMz _"BQ$WR %&?t+4*_з:2CkGy5&4߶Fd19 A~imsjT@\*H/Qk]F~hTDFl;WzF&e}ZL[AP:3m@Ȭ`7P\[[`V%=jC~xkMGuM>ܦ7.a5jks+km j>}K6u_{GTǏ#&=xuKwxVƵR[[<#\;_c= ]FOPֲx+FP۰ǿMA>(w&xֿ! i2T=o2"ɡ*㒸?&" F{PBvRDxhZ=8P1/Pvcгj_d x @Ƣ* lnηE(70p X>HC Fhhf(I7<- UԪ# r MCZ#X7;u*(NT@UErҢ-,&~<#6248T%G-p-(?Q;`w> D-^EEv{P,sC\Y9AAR7rbQZ9jР?ž`HqIo+, ђN[&_b~HF0g:=wRz~] {/G篅0^H u_9۠熋Ղ+ ̼0t.jVے X۵yBfT(~ YP{xv@>Ty>CjW[ќJ{cҭR!R2JUH`֓حgwz*Pݙ?v vwzΟ2vhU҄h}c2RM_3zEfTCk_5 aքNHmnW@e~_[KkZ^r| }|ێKc1 ޘ\eK=c!mnQMXN6u[ko4 Z3X[z-z3wl6-o=KZI; 퇻R၁3:ΠTjמ}ڿe585k|Wz۷ܢyGa,r+(3,X-_fʲõ3%hO~%n~p XN?K8Oa[.jP+:RQVvMj*$\% \2 QkV/|L޲l og`}>0kjU8UQZLÚ i,QZX@b ^Sj“zZӊ3_rkvo]em o] wbEdu* X_nEJ *J04=˖k?E}\aVP=)cyJvf{s Z5:b`yWϠK޸VJ DlOTbwcPU7xe<+|-9\ˈYp΃'*ylwkUe;{î $ۻ6;BN6ђز'*8 aad[RXk P١-*dSVUaȵDh E"ax?ݝDdyNy_!WɅ Jы$b~۾K9:͓]Q/fO~d";Ѿ>nn4<_׷g5$ĺw~^ ?ޫk ?cK+IJV3j@#ĢU\`4ʘ ]ʋC& ^31rV:[*WF:DaKo쳓sņHL{߀񹕶=ը X{#^]|^V~G//s?;UNQ6*ev{+[33bIxrhd޶O] ?+R8p~2񨷯RsT8:w>$|>2T.|s>4нcw֭C pJjpos Yqu:GxG/Ѹs6m#gI*C{vsmwN+sTpGrO915rk_G3=YF2TJ8t|T bUA[;9꧟ՇlO<.gġrɍڑnr5Lk3e5OZ?$ϟ&+oT2Oś3{>ɞ+/ܥW=I\hk{Ki^'[K,ʳY $usk$-NDǸe#rWlH8[X>ξ;nGlnfIP+*|u6I9ʼ.@ .$g_^nԔuy3ῘoP%T`e^W?nzgsI x 1_UoaU((UTA*ru#TuJ^Ktˣ 3iXùS34Q5 X\\tJMޗf{gCq1?{iы]yi-m?4L}!k\s'[mIF@9rӦВOTݺ^mX'[Ht2~nի7.ұ˩UxߩtzK%TH%VS'{2Rf.rm[mmN k"C˃]ZJ}7WxJJֶalBZmH*D~LV\:~ڗB5|QׅZgqnt4Qr!{u?e6㹫:nEU=RZ愲7۷'R7V|.%j;B•L=KWʌs{8"M-]ynW`GnlV7L}4sĕԭ2aKCM7◯b |W;ۄ#h9s'> 4\Tlçj |CXbbO^K`JL1lbyMkͲLk w܆4o.ۜs._g1Opx w  P:A?u@ P:A?u@ P:A?u@ P:A?u@ PA6-IENDB`n9nԥWU%8n KeQYʴ ъPNG  IHDR@b(gAMABO pHYs  $tEXtSoftwareQuickTime 7.1.2 (Mac OS X)3:$tIME &*:뜔 IDATx \#((h( / 00DArial009090XPX1<|D-3 000009090XPX1<| DOsaka009090XPX1<|0DTimes009090XPX1<| ` .  @n?" dd@  @@`` h[   HH  `b$khLvH>ߨ{0b$<;0_po;YF{0b$AX(x 1!L0b$> $8!^.0*0b$*x|5r>0$0b$~ x K݂O0b$Cč0:k , ,40S @8 2uʚ;J.ʚ;g4ededQ`Hppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 1___PPT10H@0 @ H8(0H  Hv___PPT9XP0`  pP`@ h___PPT2001D<4XX&___PPTMac11&@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography (   hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ? %/8Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1Ie2)1Id  1)05 R0QO- "The Expected Externality Mechanism #"$fBatch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractabilityvIZZVZVZH  VVt<ZE 'Why these characteristics are important (' *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $3Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributionsZPZZBZZtP  f/B  GGE7"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/   P` 33` Sf3f` 33g` f` www3PP` ZXdbmo` \ғq>9y`Ӣ` 3f3ff` 3f3FKf` hk]wwwfܹ` ff>>\`Y{ff` R>&- {p_/̴>?" dd@,|?" dd@   " @ ` n?" dd@   @@``PR    @ ` ` p>>   8F ( A! A! 8 8 6 "P F T Click to edit Master title style! !$ 8 0` " F RClick to edit Master text styles Second level Third level Fourth level Fifth level!     S 8 0 "`` F0___PPT106___PPT9B___PPTMac11   hnamd` Arial&Monotype Typography  T*"  8 0 "`  F0___PPT106___PPT9B___PPTMac11   hnamd` Arial&Monotype Typography  V*"  8 0 "`  F0___PPT106___PPT9B___PPTMac11   hnamd` Arial&Monotype Typography  V*" H 8 0޽h ? 3380___PPT10.2o $Blank Presentation  P o g < (  < < 6 "p F T Click to edit Master title style! ! < 0  " `   F W#Click to edit Master subtitle style$ $ < 0 "`` F0___PPT106___PPT9B___PPTMac11   hnamd` Arial&Monotype Typography  T*"  < 0 "`  F0___PPT106___PPT9B___PPTMac11   hnamd` Arial&Monotype Typography  V*"  < 0 "`  F0___PPT106___PPT9B___PPTMac11   hnamd` Arial&Monotype Typography  V*" H < 0޽h ? 3380___PPT10.2oa  \T`(    s *`J P   F P*    s *O    F R*  d  c $ ?  F  s *T  @ F RClick to edit Master text styles Second level Third level Fourth level Fifth level!     S  00[ `P  F P*    0P` `  F R*  H  0޽h ? 3380___PPT10.s$,# @( `Cp r  S в<p F l  C @~< `   F H  0޽h ? 33___PPT10u.s͸[+D=' `= @B + # p$( olppamiuk r  S *8P  F r  S *8 F H  0޽h ? 33___PPT10u.s+D=' `= @B +} # D( {5 Dl D C  $8P  F l D C p$8 F H D 0޽h ? 33___PPT10u.sG+D=' `= @B + # $$( 59xobtvcaa $r $ S &(8P  F r $ S '(8 F H $ 0޽h ? 33___PPT10u.s7<+D=' `= @B + # h(  hl h C  t1"8P  F l h C u$8 F H h 0޽h ? 3380___PPT10.s_O # pd( {5 dl d C 08P  F l d C ȱ8 F H d 0޽h ? 3380___PPT10.sМ } # @( " @l @ C `w8P  F l @ C x8 F H @ 0޽h ? 33___PPT10u.sE+D=' `= @B + #  2( ltncltnc  r  S 8P  F r  S 8P  F L  C $A depthP P  C (Apayment0@  Z  C 2Asatisfactionp P H  0޽h ? 33___PPT10u.s+D=' `= @B +# # .&0P( 4p Pl P C 8P  F l P C 8  F L P C $A front w J P C "A back. H P 0޽h ? 33___PPT10u.s~#+D=' `= @B +} # 0X(  Xl X C 8P  F l X C P8 F H X 0޽h ? 33___PPT10u.sČG+D=' `= @B + #  ,((  ! ,r , S 5(8P  F r , S p*8  F ^ , C 6Aagv expression0 l , C DA,agv payoff expression `  , 0*  `  qThe expected gain to user i is  I  , 0) @ sBecause only the left term depends on i, the expected payoff to user i is maximized by honestly revealing his type.tt> /H , 0޽h ? 33___PPT10u.sn_Y+D=' = @B + # ``( {5 `l ` C PK8P  F l ` C X8 @ F H ` 0޽h ? 3380___PPT10.s   tlP( h R  3    Fr  #  @  F  H  0޽h ? 3380___PPT10.s$,  d\  ( Kh0h R  3    Fb   C ` @  F`___PPT10@8___PPT9xp___PPTMac11tl   hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography   RHUsers receive money while waiting in the queue and pay out money at the front of the queue. Vertical segments in the payment and depth graphs correspond to jobs that did not run Horizontal segments in the payment graphs correspond to an empty batch queue (the last job to run as it empties pays nothing to others) Users whose jobs reach the front during periods of low queue depth have the greatest gain from the system. We are looking at ways to eliminate this  temporal strategization where a user strategizes about when to best enter the queue The total satisfaction graph is the Payment graph, minus the annoyances of waiting and plus the happiness of job completionN\V |[ 4  `?H  0޽h ? 3380___PPT10.s$,  x (   R  3    F~  C - @  F  H  0޽h ? 3380___PPT10.sj     ( (  (R ( 3    F  ( C p+ @<  F$H___PPT10( f___PPT9H@^___PPTMac1180   hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    Nash-equilibrium incentive compatibility : being honest about how valuable your job is to you an optimal strategy, IF everyone else does the same  A balanced budget : the Generalized Vickrey Auction lacks this characteristic. We want our mechanism to have this in order to avoid inflation, which makes it harder for participants to know how valuable the currency they have is. Inflation also disincentivises saving.  Computational Tractability : Not all ways of deploying the EEM are computationally tractable. Our formulation is tractable, however, because of the form of the user preferences. Optimally deciding whether or not to run the job at the front of the queue is possible because of the way we have users specifying their preferences. 8  K   P_Z.&H ( 0޽h ? 3380___PPT10.s(o  ( U:DHsresser:crae R  3    F  C  @  Fp___PPT10     6___PPT9     r___PPTMac11LD   hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography   For the expressions, i is the user whose payments are being computed t_i is the actual type of user i s_i is the stated type of user i t_-i is the actual type of all users other than i s_-i is the stated type of all users other than i k* is the allocation function SW_-j is the social welfare of everyone other than j, which is defined as: SUM( l != j ) of s_l ( k*(t_j, s_-j)) The last bullet point should have the nash-equilibrium caveat that the honesty maximizes his expected payoff only if everyone else is honest.. "! 2$$((M,,K0044(88<<1.MX* KH  0޽h ? 3380___PPT10.s&T  tlH(  HR H 3    Fr H # ' @  F  H H 0޽h ? 33 pP80___PPT10.sPh  tlL(  LR L 3    Fr L # ෦ @  F  H L 0޽h ? 3380___PPT10.sPh{     PT (  TR T 3    F  T C  @  Fh___PPT10H@___PPT9___PPTMac11   hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography   Each potential declared type is tested 100 times, and each test is the expected value over 100 draws. The red dots are the mean for each type. There are three other queued participants whose graphs are not shown for space reasons, but which look similar. Xf) p>{pH T 0޽h ? 3380___PPT10.sޣ  SK@\( ۂ## \R \ 3    FQ \ C o @  F@___PPT10 V___PPT980J___PPTMac11$   hnamd` Arial&Monotype Typography   For time reasons, the data supporting the claim that the simulations showed a balance budget and computational tractability were left out. &H \ 0޽h ? 3380___PPT10.s1 dxp^RЀ3ÿ lHbP  @AL G@;b `B&V+~m}? )a]<57  =oA~p{Eoҥ ڪ6dF\ZQE ,-͕K(!֖*.wqN[+<LŊyމˆZ|Zz]zv* oAXZNAƙhg`| ]. @t&ڮ9)e],Zqf8sK`TdN<4G,i͔Fty6O`7fDu+OVřF/uɄ?ϻAgk$53I 7g;pǙ605^@7z'nlXeIh4믿43]O ̆Y=n&o Z GH% j3:tWh8E&Mah߬<6|4QRQM| @_Ljϻ~>aӊJN;+bkӘ3Kځ3maCs:90̳A+}ǎ+]]]vڑ{ O>O  &ϾrʨQ}6 nݺu֣<|+00V"HĮizey? >\gӈѣ֭[nnngϞuvvVCA~7!,ǮF?釘Ͼp} y_lܸիo%XVw>&w>}3o޼ݻ۷ɣ7pHQQѹsmhhVK#bqǎӧOg]nB~+A~_Ӫ&MgҚJKK:ujFF6|=o%oBߗFVL /.[n={sΜ͒+J%.\ 삂_|Q:ѣGW^y%88u- Ԥ3;;;{vss300h۷G7gV H$.]gvzz<ǎ3:~u]fbbº݁V xѕ+W}E{{{yf={d]Z,^ì o%oP-?s̞0aB^XWvꫯ  pl߾]d͚5(++SLVUU mȐ!K"4P^yqƅZo%%%KHNNg;wӪ1uiZFNttɓǏϺ^=ex*bX1J^^vqqcm7,_<++KcCIX VSSsEyfqwwg6iUZreyy#GXao%oJUqttgq㌌XWy+u͛ǺB~+A~Gedd3;%%WL:_z饴Jzɓ4/l>onnκ4yС/{J&**M&m=VSx0օp[ @ܹsgѢE山Æ c]_?'Z8>ᑖfO>bܹg]A :t?stt4X2k׮-((8qB7葸8''#G c]4غu[ǺtAyy%K~駘@SǏOII6B݌{:W_}@իodoom6ooﺺ:ֵ}-,,KF+V̌vwwg]bƌ}B8w3膄v?4ޜ_f].[2N>/[HjHiD`0y}J>!f].ˏ?/dggcxwkkkoκ1`]@["p\od0zB]w.6l w:HͯpsV!իn9rdWmA?CBB:曬ڵkllQ&ML~.}Us@yU&`>|8múдDFFΚ5+##gϞ䷱_Ttx=߀$o[<%RIx̛<IKwwC"lڴi߾}v߬fN>ֵpg]y/__јIp\T"%fDTTp8@ʕ+Zsڵs4(++o߾W_g] pf5 IιS#*6$^ocBz4$J7o/[.!! TϞ=-[?8^] wfܳ] 8[%[{74dGp2/$ńuG}Eyͺ"F_BGXע4ZEü l'[i^30 0$v{kl=|?ܾ}{xxEXڈ~DO4iĉ6l`] K^MiqZ V>U\^XVEu #ə7o{9*))qqqIHH3f Zѽ7= ۲eҥKY@; ++~c] o%oϧmCvDhe˖UUU} aC_B?zv=f̘Yf%%%!|GW\e]B jq HRv2u9IYYY'NLOO4hZ4 70w^77_… oh7GG{o ZoL .g]W_}~ܼy3B4 )W_}X2c v_~͚5֭ڵ+r@GF/|Yע9o%h|jcǎ+@w,Z~Ҷ{Ĉ]sԩ%KdeeՋu-V@38E@@;whκ A~+A~t_~eEEE4X͍… Yע ȑ#4/_u޽{||uXע5W/7kqՋZaS/[oLL̨QXzd޽t钡!ZK[ן3B25epܓB$Գ qǏ_|7mԭ[7y7?օ?ό 7ϐDYxCD%9B!gT0 Zv޽ٳg_>!!a۶mo`>=wBc 3Nlg tQ-q_UrRj%666YYYGf]޽{믬kaёf?Dv l-b0m+/^2qķzk…k%Za;/yﷅ5_h|io)[{:^s1cR٩SzEn7h-[ܹsO>a]3l_+=fJ*?"#)Nj+'ӫ'$Գ_hիW;w|>u9M;… /Z:f=94)¯ ?Τ)hlӶ@$!A; :?e] ֈFZbjn^f[)J,[_m_~ez׮] dZ65c`bޔ/ѦMQpy<^]] ;y~˺MSk-s<ڌ>>4?SօtTf̨w1-!Hl'ڵk٬|NNN'O6mZԋIT/&Ɏ([E[{^ׯrȑ#SNuss߿?rԈ-/9 4A*Ҷ{ҤI_5t{@@[o#ֵ6y!qTnj9o<hix Q#Fb5^XvJk XezÇ녅.@n߾=ro͍u-gocsNKm>l O;IkyjZy>.|>13k>joʕ{QhڣGvo޼yҥ:3$Ӛ1cӧ~Co;$/&,3{l5+((?>A.1]9>>ۛu-k<77~qc_& ^ ;22244G 1eʔ˗/ڲ)hoBfx+6 3JBlȴnjzT7[n-XCe]vٱc/\еkWֵU /28ǟO}mIQ\駟oKIIAxի{eօt&6ۿ:o3",`ekW?uv޾?/\MaÆ.@K^6&&iƍc]N`W|<ˣdIjcfN?~8#G8q?hY~8  ]ֵt&q!_y#A'^z ,%''.M*2WMp-j6&)+_v7y ;;MIWW9.4]ohE;<<~\~Z]'KKÇ[YYYi6i<Ba3n( ‰Ppbi4{_?eЌɓ'{{{?~u-$9h} mQ^4y˔pQdE][J' M6h'۶m=o<$7߿Ϗu-OM];mބTo[%dͻc(bFDIE5;oM[_Ct_~Esϟ?ƍc]^֭[||q^z%{{{<&xoIpor33o9s檥~|" NJ~=];@H*VhÇӗ1$۶mc]auX.q |Ц bsosh3=0gX$Ip1-|%7􍼐|?Z&78| ˺"o4hǎ ybӧp19.ɍ m}7";~LZ_!QnMo-QZZ.O$7sxʔ)SXVLAHfDavmEc;ZwMns̟?=7V9Y,--Y"W7u73ъh"i} $70a}{| Zڄk<]4ihrZjĈ4w؁N6m*//߳gBڄk5Gm D`otՀցw֟& =… y6+((=ztrr2Xl;s.&QQ=j!A3JJJk-ZDʊuEVW\12RS/&ڹٸ̬*+kBܱErpϩS֮]u--aKUfGܴ^n*++y<^ddԩSYՠ͎0xǩ5t^#pA4W\`dd"8ȑ#xii)ZTbgob7?(b7:Py":XɽuVs16mpٳg١aZ!5oO|yd20bߛ Pۊu?΢Hn???Z2tN0aڴik׮km~+Ȃ?}mbH?'OD~CݹsG 󊊊N>=rH ?nnm5,*%t\\ ^@^bC=~> lmm3k֬)-./L>7[yx/Eh{Ng9}ƘkJKwZX̨`RMe˖x<333so۶ rN׮]XҀkG{7bhhȨwիW~~> qll].]ĺ6]z2;|I#j} B4,YlaaA8q||[6& qKCG]ջw_BLe|Jևo˖-ϧͺ"*'' 6̚5+55k׮Lj`s87tDe-b5QՋԲ ɽf$7t^{}ӦM-=~nUARPrZJ&ad<7!܋/vvvӧi#S|=|{6hI3ZS^Z&%ɽe˖jWFl::ujɒ%YYYz{־W߈Q9^[HnЌUV=zsVtQEȲ;#[=cǐ13gxZ[RȰW{P5nu떟K~ h#@3wnݺ^I%q{yn%tMn'N=7MnssL /l޼y֬Y?fT&n`7zvIŚ+tMn__QF7XYlY7lؠGcREsk^ ͛7'M?sidr5^h+M޽}||ݻcƊKFMhĽF4~~g &M4gΜ j4xXMc8BBa-ⒸdφްL3|w{&#µM-[`gB/K'$:4;Sd TO'{gffƺ"kc`1?զ~?n B<ۆ9ۜ$+R+ܜUnzF٩U̙3ghMu!3m4`] wNNHE`}-Ș|H=n[HP>w@{p&<ٛfUZXVEQC8߭03d]pf5P@~p{܃]i;vLqyƌ +h}︸8e7p>w{܃7  =oA~p{܃}Tqa%]RRºR[[#U}7M~˗/. @t?Xנ9]t?)kLMMMZZ}wޯڻwoֵpRjkkuƺCG@ֵp}ҷ0}ҷ0Z8 4D:9RbFDIE55:BxM[_]3_XQ}| i%-|~[% ėnzqiPd^HKs; &·Dx>eWsY7e07@:v5zP"iq-[O>i_|յN9;;gffv:Ȩ#[o}$OOώ0x`=tDyyeG7HHHh'&&pf[XYYu:ή#0q;gΜ`ll\]]ݑ{k׮999u{]ti\*hcW5I>8Do}Tpm27W p{܃7 Ǐ5ڤ RqECcFT T\^^%5004gWke&K,T R9 Z 0)ψ=~TJQ[UT ` *Dȓ,)`47zv^G /uQ[ާތf~}bjQd [DwxUnlu`AN@UEr'D8F|6b$'""EEկ H)ȦTOEQ~TjəNiu,:OG4仹u]4U{hLUlI]%ljUhmU8u}C_Xy z(Qb-T Eoc" oۘFK ]:%f:0.W|@@v 07%A߿#zUR5,ȩA'5ՖYxBli:&_iN 93O qyj !64M6}/}X>5ebh5SOE~p?=oA~p{܃?Dv~KſOJ*tc^kEnv8@ ŢSۯxDq?+h|x|tD=bDHn$so~l̈}Evڻ474{`08A(aC+*6467iR*ץÓ& 1457Aq ԟ:I,]qq*.sv=qU1酙1ot140 ?v ɔ7F2Zvx@(|`DB,%o6@M\eKo |\fLgltlF iTJMH;į_ L*NHH]Y^A3Ro׾81pKG6A~@Y;>ȍ DןjbR(mf=~$19Q@&ؙa6m\L]J5⫙ei~Kϯg'}j2f8yAI٩vsUٿ@fHS\ЊKKxSmMe߷_0Ʋ;0WF,|7.?m2>}:&ˋoݭ4afaneinĺC~pp{܃7  =oA~p{܃7  =oA~p{܃7  =oA~p{܃7  =oA~p{܃7?R0tƊRIENDB`0DArial009090HpPHp1<|D-3 000009090HpPHp1<| DOsaka009090HpPHp1<|0DTimes009090HpPHp1<| ` .  @n?" dd@  @@`` h[    HH  `b$khLvH>ߨ{0b$<;0_po;YF{0b$AX(x 1!L0b$> $8!^.0*0b$KeQYʴ ъ970$0b$~ x K݂O0b$Cč0:k , ,40S @8 2uʚ;J.ʚ;g4ededQ``ppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 1___PPT10H@0 @ H8(0H  Hv___PPT9XP0`  pP`@ h___PPT2001D<4XH&___PPTMac11&@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography (   hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ? %/8Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1Ie2)1Id  1)05 R0QO- "The Expected Externality Mechanism #"$fBatch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractabilityvIZZVZVZH  VVt<ZE 'Why these characteristics are important (' *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $3Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributionsZPZZBZZtP  f/B  GGE7"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/   #  "( ltncltnc  J  C "A soloP %r  S 8P  F r  S 8P  F L  C $A depth!P  P  C (ApaymentW @ H  0޽h ? 33___PPT10u.s+D=' P= @B +r_acpnn( / 00DArial009090HpPHp1<|D-3 000009090HpPHp1<| DOsaka009090HpPHp1<|0DTimes009090HpPHp1<| ` .  @n?" dd@  @@`` td    HH  `b$khLvH>ߨ{0b$<;0_po;YF{0b$AX(x 1!L0b$> $8!^.0*0b$KeQYʴ ъ970$0b$~ x K݂O0b$Cč0:k , ,40S @8 5uʚ;J.ʚ;g4ededQ`Hppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 40___PPT10H@H @ H8(0       H___PPT90  pP`     @ h___PPT2001D<4XH(___PPTMac11(@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography 8"   hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnatype Typography    hnamd` Arial&Monotype Typography    hnamd` md` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography p   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ? %18Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations and scheduling decisions to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1ZIZeZ2Z)ZZ1Id  1)05 R0jO- "The Expected Externality Mechanism #"$8Batch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractability We plan to investigate similar solutions for reservation environments, where there are no externalitiesIZZVZVZiZH  VVg  <ZDh 'Why these characteristics are important (' *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributions Providing these features in reservation systems is easier, as there are no external relationships between consumers of different time slots We plan to investigate similar solutions in a reservation environmentZPZZBZZZFZt  Of/  $$((A,,004488F<<GGE7;"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/P     } # D( {5 Dl D C  $8P  F l D C p$8 F H D 0޽h ? 33___PPT10u.sG+D=' P= @B + # $$( 59xobtvcaa $r $ S &(8P  F r $ S '(8 F H $ 0޽h ? 33___PPT10u.s7<+D=' P= @B +  # h(  hl h C  t1"8P  F l h C u$8 F H h 0޽h ? 3380___PPT10.s_O  # pd( {5 dl d C 08P  F l d C ȱ8 F H d 0޽h ? 3380___PPT10.sМ } # 0X(  Xl X C 8P  F l X C P8 F H X 0޽h ? 33___PPT10u.sČG+D=' P= @B +  # ``( {5 `l ` C PK8P  F l ` C X8 @ F H ` 0޽h ? 3380___PPT10.s   " tll(  lR l 3    Fr l # |, @  F  H l 0޽h ? 3380___PPT10.sVlh  " tlp(  pR p 3    Fr p # , @  F  H p 0޽h ? 3380___PPT10.sVlh  " tlt(  q,`z, tR t 3    Fr t # Ж, @  F  H t 0޽h ? 3380___PPT10.sVlhr@c i0I)0 yc n( / 00DArial009090HpPHp1<|D-3 000009090HpPHp1<| DOsaka009090HpPHp1<|0DTimes009090HpPHp1<| ` .  @n?" dd@  @@`` td    HH  `b$khLvH>ߨ{0b$<;0_po;YF{0b$AX(x 1!L0b$> $8!^.0*0b$KeQYʴ ъ970$0b$~ x K݂O0b$Cč0:k , ,40S @8 5uʚ;J.ʚ;g4ededQ`Hppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 40___PPT10H@H @ H8(0       H___PPT90  pP`     @ h___PPT2001D<4XH(___PPTMac11(@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography 8"   hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography p   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ? %18Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations and scheduling decisions to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1ZIZeZ2Z)ZZ1Id  1)05 R0jO- "The Expected Externality Mechanism #"$8Batch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractability We plan to investigate similar solutions for reservation environments, where there are no externalitiesIZZVZVZiZH  VVg  <ZDh 'Why These Characteristics are Important ('  *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributions Providing these features in reservation systems is easier, as there are no external relationships between consumers of different time slots We plan to investigate similar solutions in a reservation environmentZPZZBZZZFZt  Of/  $$((A,,004488F<<GGE7;"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/P     }  # h(  hl h C  t1"8P  F l h C u$8 F H h 0޽h ? 33___PPT10u.s_O+D=' P= @B +rnU JW pn( / 00DArial009090HpPHp1<|D-3 000009090HpPHp1<| DOsaka009090HpPHp1<|0DTimes009090HpPHp1<| ` .  @n?" dd@  @@`` td    HH  `b$khLvH>ߨ{0b$<;0_po;YF{0b$AX(x 1!L0b$> $8!^.0*0b$KeQYʴ ъ970$0b$~ x K݂O0b$Cč0:k , ,40S @8 -6uʚ;J.ʚ;g4ededQ`Hppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 40___PPT10H@H @ H8(0       H___PPT90  pP`     @ h___PPT2001D<4XH )___PPTMac11(@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography 8"   hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography p   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ( ʚ;&)ʚ;dd? %18Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations and scheduling decisions to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1ZIZeZ2Z)ZZ1Id  1)05 R0jO- "The Expected Externality Mechanism #"$8Batch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractability We plan to investigate similar solutions for reservation environments, where there are no externalitiesIZZVZVZiZH  VVg  <ZDh 'Why These Characteristics are Important ('  *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributions Providing these features in reservation systems is easier, as there are no external relationships between consumers of different time slots We plan to investigate similar solutions in a reservation environmentZPZZBZZZFZt  Of/  $$((A,,004488F<<GGE7;"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/P     rWW /n( / 00DArial00BBHpPHp1<|D-3 00000BBHpPHp1<| DOsaka00BBHpPHp1<RdO)@Pictures<PowerPoint Document(Q'SummaryInformation(t     '+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwy)z{|}~*     !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~ eB  !"#$%&'()*+,-./0123456789:;<=>?@ACDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdfghijklmnopqrstuvwxyz{|}~DocumentSummaryInformation8Current UserM oelbelhnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography 8"   hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography p   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ( ʚ;&)ʚ;dd? %18Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations and scheduling decisions to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1ZIZeZ2Z)ZZ1Id  1)05 R0jO- "The Expected Externality Mechanism #"$8Batch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractability We plan to investigate similar solutions for reservation environments, where there are no externalitiesIZZVZVZiZH  VVg  <ZDh 'Why These Characteristics are Important ('  *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributions Providing these features in reservation systems is easier, as there are no external relationships between consumers of different time slots We plan to investigate similar solutions in a reservation environmentZPZZBZZZFZt  Of/  $$((A,,004488F<<GGE7;"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/P      6    "( ltncltnc  L  C $A depthP `J  C "A soloP % r  S 8P   r  S p8P   P  C (ApaymentW @ H  0޽h ? 33___PPT10u.s+D=' P= @B +r59 Fn( / 00DArial00BBHpPHp1<|D-3 00000BBHpPHp1<| DOsaka00BBHpPHp1<|0DTimes00BBHpPHp1<| ` .  @n?" dd@  @@`` |d    HH   `b$khLvH>ߨ{@b$<;0_po;YF{@b$AX(x 1!L@b$> $8!^.0*@b$KeQYʴ ъ97@$@b$~ x K݂O@b$Cč0:k , ,4@S @8 -6uʚ;J.ʚ;g4ededQ``ppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 40___PPT10H@H @ H8(0       H___PPT90  pP`     @ h___PPT2001D<4XH )___PPTMac11(@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography 8"   hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography p   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ( ʚ;&)ʚ;dd? %18Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage #The Expected Externality Mechanism(Why These Characteristics are ImportantPreference Specification The EEM in a Queued EnvironmentDemonstration&Demonstrating Incentive CompatibilityIn Conclusion#The Expected Externality MechanismDetermining Job Value  Fonts UsedDesign Template Slide Titles '_-'{Charles KoelbelCharles Kin order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations and scheduling decisions to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1ZIZeZ2Z)ZZ1Id  1)05 R0jO- "The Expected Externality Mechanism #"$8Batch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractability We plan to investigate similar solutions for reservation environments, where there are no externalitiesIZZVZVZiZH  VVg  <ZDh 'Why These Characteristics are Important ('  *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributions Providing these features in reservation systems is easier, as there are no external relationships between consumers of different time slots We plan to investigate similar solutions in a reservation environmentZPZZBZZZFZt  Of/  $$((A,,004488F<<GGE7;"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/P     rV929 n( / 00DArial00BBHpPHp1<|D-3 00000BBHpPHp1<| DOsaka00BBHpPHp1<|0DTimes00BBHpPHp1<| ` .  @n?" dd@  @@`` |d    HH   `b$khLvH>ߨ{@b$<;0_po;YF{@b$AX(x 1!L@b$> $8!^.0*@b$KeQYʴ ъ97@$@b$~ x K݂O@b$Cč0:k , ,4@S @8 -6uʚ;J.ʚ;g4ededQ`Hppp@ <4!d!d@Egʚ;<4dddd@Egʚ; 40___PPT10H@H @ H8(0       H___PPT90  pP`     @ h___PPT2001D<4XH )___PPTMac11(@f   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography 8"   hnamd` Arial&Monotype Typography  d   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  @   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  4   hnamd` Arial&Monotype Typography ,   hnamd` Arial&Monotype Typography D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography X   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography p   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography  (namd Osaka  D   hnamd` Arial&Monotype Typography    hnamd` Arial&Monotype Typography ( ʚ;&)ʚ;dd? %18Eliciting Accurate Job Value Information in Batch Queues98$ /CAndrew Mutz and Rich Wolski University of California, Santa Barbara,D', (  Introduction  Not all jobs are of equal value to users First-come-first-served mechanisms that ignore job value can lead to less important work being done before more important work Without properly designed mechanisms, asking a user the value of a job can yield unreliable answers Users are motivated to overstate job importance The field of economics known as  Mechanism Design has solutions to these problems that we can leverage in order to elicit accurate information from usersZ0ZZ)~  d0  >A,Using Economic MechanismsCurrency is distributed by the resource provider Can be distributed non-uniformly, to give some users scheduling benefits Users will enter a queue and jobs will be considered for running in a first-come-first-serve manner Bids will be placed, low value jobs will not run Why not use a simple eBay-style auction? In order for the currency allocations and scheduling decisions to be meaningful, the bids must reflect the true value of the work being performed The decision to run affects all participants1ZIZeZ2Z)ZZ1Id  1)05 R0jO- "The Expected Externality Mechanism #"$^Batch Queues have characteristics economists refer to as  externalities The decision to run the job at the front of the queue affects everyone, not just the job owner and the resource provider These externalities restrict the space of mechanisms that we can use The Expected Externality Mechanism (K.J. Arrow, 1979) can deal with these externalities while providing: Nash-equilibrium incentive compatibility A balanced budget Computational tractability We plan to investigate similar solutions for reservation environments, where there are no externalitiesIZZiZVZiZH  iVg  <mDh 'Why These Characteristics are Important (' *Incentive Compatibility Users optimal bidding strategy is to be completely honest For the scheduling decisions to be useful, accurate value information is necessary Strategization by users is inefficient Balanced Budget Imbalanced mechanisms cause inflation/deflation, which is a headache for users and administrators Computational Tractability We want mechanisms that can be deployed in the real worldZZZbZZ:Z;S'  b:bHU Preference Specification $When a user submits a job, he is asked to specify his preferences In terms of currency, how valuable is the job? i.e. this job is worth 10 of my 100 credits In terms of currency, how much do you dislike waiting in the queue? i.e. for every hour you can knock off my queue residency, I will pay 2 credits Why this preference specification? The theoretical properties of the Expected Externality Mechanism rely on an optimal allocation function Very expressive preference specifications are computationally hard to optimally solveBZ/Z,ZDZOZ$ZZB/,D  O$hU   )Q:"IDUFThe EEM in a Queued Environment $Whenever the resource becomes available, the allocation mechanism compares the stated value of the job at the front of the queue the aggregate dislike of waiting of everyone else If the job s value exceeds the imposition on the jobs behind it, the job runs. Otherwise, it does not. Once the allocation has been made, the participants are charged according to the EEM payment scheme The EEM uses a statistical distribution of participants preferences Users are motivated to maximize expected social welfareLjhd}Jjh  d}A5\u Demonstration  $The EEM payment expression is cumbersome, so here is an example Each time step has a 25% chance of adding four jobs. Job values distributed uniformly between 5 and 100 Dislike of waiting distributed between 2 and 4Nvb@53  /Z %Demonstrating Incentive Compatibility &%$ In a single step, how is the expected payoff of a participant affected by their announcement? Single instance with the job value drawn from (5, 350) and the dislike of waiting drawn from (20, 100) If the actual type is a maximum, then honesty is the best strategyP ^g  BP]MC In Conclusion  $Our tests indicate that the EEM makes honest preference revelation an optimal (Nash-equilibrium) strategy for users. Future Directions Better understand the strategic dynamic of when users choose to enter the queue Currently, entering the queue as it is emptying yields higher payments than entering as it is growing Can we modify the mechanism to ameliorate this? Build and deploy the mechanism in an environment with real users Compare the performance of the system to existing systems Use the information we gather to build realistic preference distributions Providing these features in reservation systems is easier, as there are no external relationships between consumers of different time slots We plan to investigate similar solutions in a reservation environmentZPZZBZZZFZt  Of/  $$((A,,004488F<<GGE7;"The Expected Externality Mechanism##$"The mechanism relies on knowing the statistical distribution of the participants preferences The payment of user i is 6y^ PM Determining Job Value $, Assume the user plans to use his credits for two jobs, A and B W.L.O.G., assume the user knows V(A) > V(B), but the user does not know V(A) or V(B). V(A) + V(B) = total credits An  ordinal relationship can be converted into a  cardinal relationship through a series of questions Which would the user prefer: completing A with probability 0.5, or completing B with probability 1.0? Depending on the answer to this, adjust the probability 0.5 until the user is indifferent If the user is rational, he is making decisions to maximize P(A)*V(A) + P(B)*V(B). Combining the above equation with P(A)*V(A) = 1.0*V(B) allows us to solve for V(A) and V(B)ZZiZZTZ\Z?V  ifZT\  >97gp%/P      6  $$( 59xobtvcaa $r $ S 78P   r $ S 78  H $ 0޽h ? 33___PPT10u.s7<+D=' P= @B +Z  "   (j (  (R ( 3      ( C Ќ; @X  @P___PPT100(v___PPT9XPb___PPTMac11<4   hnamd` Arial&Monotype Typography     hnamd` Arial&Monotype Typography   0V Nash-equilibrium incentive compatibility : being honest about how valuable your job is to you an optimal strategy, IF everyone else does the same  A balanced budget : the Generalized Vickrey Auction lacks this characteristic. We want our mechanism to have this in order to avoid inflation, which makes it harder for participants to know how valuable the currency they have is. Inflation also disincentivises saving.  Computational Tractability : Not all ways of deploying the EEM are computationally tractable. Our formulation is tractable, however, because of the form of the user preferences. Optimally deciding whether or not to run the job at the front of the queue is possible because of the way we have users specifying their preferences. Arrow, K. J. (1979):  The Property Rights Doctrine and Demand Revelation under Incomplete Information, in Economics and Human Welfare, ed. by M. Boskin, New York: Academic Press, 23-39B  K   t_Z.$"H ( 0޽h ? 3380___PPT10.s(orR .' PRdO)tyPictures<PowerPoint Document(Q'SummaryInformation(t     +,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwyz{|}~DocumentSummaryInformation8Current User) #The Expected Externality Mechanism(Why These Characteristics are ImportantPreference Specification The EEM in a Queued EnvironmentDemonstration&Demonstrating Incentive CompatibilityIn Conclusion#The Expected Externality MechanismDetermining Job Value  Fonts UsedDesign Template Slide Titles _-'x xx x5     { !"#$%&'()*+,-./012346789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz|}~5     x{ !"#$%&'()*+,-./012346789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz|}~