Z1" x= Ox%DA*
objective/klĽ C !' F x%DA*
objective/kl_coefO>^R' F x%DA*
objective/entropyzB:/ m] P x%DA*"
ppo/mean_non_score_rewardML% 6 x%DA*
ppo/mean_scores0<)$ B+M y%DA*
ppo/std_scores_< n- $ B+M A}%DA*
ppo/loss/valuebA$ B+M l}%DA*
ppo/loss/total?%( pJ }%DA*
ppo/policy/entropyJ?j) 7_ }%DA*
ppo/policy/approxkl@ѽ) 7_ }%DA*
ppo/policy/policyklid?r) 7_ }%DA*
ppo/policy/clipfrac33>EȠ0 _ }%DA*#
!
ppo/policy/advantages_mean3w& sO }%DA*
ppo/returns/mean@7/Uى% 6 }%DA*
ppo/returns/varΊ.B # wC a~%DA*
ppo/val/vpred - 7# wC ~%DA*
ppo/val/errorDAnj& sO %DA*
ppo/val/clipfracff>A" x= %DA*
ppo/val/meanS! { %DA*
ppo/val/vargAZױ+ K %DA*
ppo/val/var_explained >+&' F I%DA*
ppo/learning_rateo:"+ K %DA*
time/ppo/forward_pass%>y J. W %DA*!
time/ppo/compute_rewards @:f1 D%DA*$
"
time/ppo/compute_advantages HI;ۚ, E %DA*
time/ppo/optimize_stepCd?T) 7_ %DA*
time/ppo/calc_stats0 >'%$ B+M 3%DA*
time/ppo/totalx?c % 6 %DA*
env/reward_mean0< $ B+M ׂ%DA*
env/reward_std_<@/t" x= qCEA*
objective/klB' F CEA*
objective/kl_coefO>r ~E' F CEA*
objective/entropyrOB
t/ m] P CEA*"
ppo/mean_non_score_rewardNS % 6 CEA*
ppo/mean_scores:R= $ B+M *CEA*
ppo/std_scores¿>2- !S) 7_ CEA*
ppo/policy/policyklה>yE) 7_ CEA*
ppo/policy/clipfrac33>0 _ 'CEA*#
!
ppo/policy/advantages_mean33s283& sO 9CEA*
ppo/returns/meansS<4% 6 JCEA*
ppo/returns/vargB~_# wC ZCEA*
ppo/val/vpredlbP# wC jCEA*
ppo/val/error~#ApuD& sO CEA*
ppo/val/clipfrac>4gQ" x= CEA*
ppo/val/meanX6! { CEA*
ppo/val/vardA + K CEA*
ppo/val/var_explainedJ:?d' F CEA*
ppo/learning_rateo: 8h+ K Z CEA*
time/ppo/forward_pass)>-U. W CEA*!
time/ppo/compute_rewards :o1 CEA*$
"
time/ppo/compute_advantages K;y_, E t CEA*
time/ppo/optimize_step\zd?RSh) 7_ CEA*
time/ppo/calc_statsp >h$ B+M "!CEA*
time/ppo/total?+% 6 y!CEA*
env/reward_mean:R= L$ B+M !CEA*
env/reward_std¿>α" x= XIA*
objective/klAB;%' F lXIA*
objective/kl_coefO>V' F XIA*
objective/entropyzB/ m] P XIA*"
ppo/mean_non_score_reward֝9'% 6 XIA*
ppo/mean_scores =$ B+M XIA*
ppo/std_scores)=Dr- Kͫ) 7_ wXIA*
ppo/policy/policyklk7=گV) 7_ XIA*
ppo/policy/clipfrac >Vbq0 _ =XIA*#
!
ppo/policy/advantages_mean At& sO XIA*
ppo/returns/meanUg 'Ƿ% 6 XIA*
ppo/returns/varB# wC XXIA*
ppo/val/vpred># wC XIA*
ppo/val/error߀A1& sO
XIA*
ppo/val/clipfracff>tF" x= aXIA*
ppo/val/mean#i`! { XIA*
ppo/val/varAb9+ K XIA*
ppo/val/var_explained?| ' F tXIA*
ppo/learning_rateo:kɼ+ K XIA*
time/ppo/forward_passo1>%. W *XIA*!
time/ppo/compute_rewards X:k1 XIA*$
"
time/ppo/compute_advantages C;
^0, E XIA*
time/ppo/optimize_step
l?%`) 7_ 3XIA*
time/ppo/calc_statsP $>BF$ B+M XIA*
time/ppo/totalP?;Wy% 6 XIA*
env/reward_mean =C|z$ B+M AXIA*
env/reward_std)= K" x= JA*
objective/kl"uBM' F cJA*
objective/kl_coef
O>7' F JA*
objective/entropy6sBԵ*/ m] P JA*"
ppo/mean_non_score_reward.% 6 JA*
ppo/mean_scoress='1y$ B+M JA*
ppo/std_scoresM=1PG- i3W1) 7_ uJA*
ppo/policy/policykl?L>) 7_ JA*
ppo/policy/clipfrac>Ha<0 _ >JA*#
!
ppo/policy/advantages_mean 3vN& sO JA*
ppo/returns/meanzv V% 6 JA*
ppo/returns/varWBo_]# wC kJA*
ppo/val/vpredR^7# wC JA*
ppo/val/error*A0=& sO %JA*
ppo/val/clipfrac? b" x= JA*
ppo/val/mean@Ȋ! { JA*
ppo/val/varh]AaT^+ K OJA*
ppo/val/var_explainedV>5?!' F JA*
ppo/learning_rateo:uH+ K =JA*
time/ppo/forward_pass=0>{. W JA*!
time/ppo/compute_rewards P:ǡ(1 JA*$
"
time/ppo/compute_advantages TE;* , E t JA*
time/ppo/optimize_stepl?k) 7_ JA*
time/ppo/calc_stats#>+P$ B+M +
JA*
time/ppo/totalꄡ?C\S% 6
JA*
env/reward_means="2 $ B+M
JA*
env/reward_stdM=o2?" x= KA *
objective/klB#' F KA *
objective/kl_coef P>a ' F KA *
objective/entropyVBt/ m] P $KA *"
ppo/mean_non_score_rewardRVa% 6 8KA *
ppo/mean_scoresu=
$ B+M LKA *
ppo/std_scores>`- ( pJ læKA *
ppo/policy/entropy?@4) 7_ æKA *
ppo/policy/approxklt>˪$) 7_ æKA *
ppo/policy/policyklq=.) 7_ æKA *
ppo/policy/clipfrac43]>JD
0 _ æKA *#
!
ppo/policy/advantages_mean/Y& sO æKA *
ppo/returns/mean{dT% 6 æKA *
ppo/returns/varqBf Ԃ# wC æKA *
ppo/val/vpred C# wC ĦKA *
ppo/val/error[=@Ϻ܌& sO cĦKA *
ppo/val/clipfracff>" x= ĦKA *
ppo/val/mean^$ ! { ŦKA *
ppo/val/varTu BB;E+ K mŦKA *
ppo/val/var_explained0L??R' F ŦKA *
ppo/learning_rateo:2+ K ƦKA *
time/ppo/forward_pass@q2>6. W qƦKA *!
time/ppo/compute_rewards :1 ƦKA *$
"
time/ppo/compute_advantages D;!`!, E ǦKA *
time/ppo/optimize_step$l?cD) 7_ sǦKA *
time/ppo/calc_stats$>7qa$ B+M ǦKA *
time/ppo/total?MP% 6 ȦKA *
env/reward_meanu=8t$ B+M qȦKA *
env/reward_std>F\" x= &LA *
objective/kl B>' F *'LA *
objective/kl_coef5
' F G'LA *
objective/entropylB"/ m] P ]'LA *"
ppo/mean_non_score_reward bL(% 6 r'LA *
ppo/mean_scoresּA=c.$ B+M 'LA *
ppo/std_scores 9y=MV=- 3E( pJ S,LA *
ppo/policy/entropy!@g) 7_ k,LA *
ppo/policy/approxklEO>z#) 7_ ~,LA *
ppo/policy/policykl>CL() 7_ ,LA *
ppo/policy/clipfracff>R0 _ ,LA *#
!
ppo/policy/advantages_mean ss& sO ,LA *
ppo/returns/meanu% 6 1-LA *
ppo/returns/varBaJ# wC 1LA *
ppo/val/vpredNnQr# wC 1LA *
ppo/val/error@Cx& sO 1LA *
ppo/val/clipfrac43>P{D$" x= 1LA *
ppo/val/meand+/!o! { 1LA *
ppo/val/var4QA'8ˢ+ K 2LA *
ppo/val/var_explained& Y?L%' F 2LA *
ppo/learning_rateo:Q&+ K /2LA *
time/ppo/forward_pass />8V. W ?2LA *!
time/ppo/compute_rewards :{1 4LA *$
"
time/ppo/compute_advantages K;ye, E 4LA *
time/ppo/optimize_stepl?ւ\)) 7_ 4LA *
time/ppo/calc_stats#>>$ B+M 5LA *
time/ppo/totalJ?G% 6 6LA *
env/reward_meanּA=#$ B+M ~6LA *
env/reward_std 9y=qQ" x= MA *
objective/klB;N' F FMA *
objective/kl_coefS^P>٩O' F bMA *
objective/entropy`B7d[/ m] P uMA *"
ppo/mean_non_score_rewardvO|~em% 6 MA *
ppo/mean_scores^ =O`$ B+M MA *
ppo/std_scores)F=-- O( pJ MA *
ppo/policy/entropy@4) 7_ MA *
ppo/policy/approxklj=2) 7_ #MA *
ppo/policy/policyklp=P) 7_ RMA *
ppo/policy/clipfrac @>>0 _ iMA *#
!
ppo/policy/advantages_mean @& sO MA *
ppo/returns/mean /|#% 6 9MA *
ppo/returns/var A@C# wC PMA *
ppo/val/vpreduWxl# wC aMA *
ppo/val/errorp-@h?& sO qMA *
ppo/val/clipfrac̃>v" x= MA *
ppo/val/meanf !! { MA *
ppo/val/var. W MA *!
time/ppo/compute_rewards 8:1 oMA *$
"
time/ppo/compute_advantages @C;d, E µMA *
time/ppo/optimize_step@l?yK) 7_ MA *
time/ppo/calc_stats@#><$ B+M vMA *
time/ppo/total¡?â7% 6 MA *
env/reward_mean^ =钥$ B+M MA *
env/reward_std)F=$" x= OA *
objective/klНBWb' F OA *
objective/entropy?2BO/ m] P OA *"
ppo/mean_non_score_rewardMSW% 6 OA *
ppo/mean_scoresp=q$ B+M OA *
ppo/std_scores 9>}
N- q( pJ OA *
ppo/policy/entropy @ur ) 7_ OA *
ppo/policy/approxkl(Z>!) 7_ OA *
ppo/policy/policykl=+4) 7_ OA *
ppo/policy/clipfracff$>)0 _ & OA *#
!
ppo/policy/advantages_meanffn}!& sO 9 OA *
ppo/returns/mean% 6 J OA *
ppo/returns/varFA## wC Z OA *
ppo/val/vpredS]o# wC k OA *
ppo/val/error `@ :`& sO z OA *
ppo/val/clipfrac23>g" x= OA *
ppo/val/meanݧЭL! { OA *
ppo/val/var5RAg+ K OA *
ppo/val/var_explainedK?8' F OA *
ppo/learning_rateo:+ K OA *
time/ppo/forward_pass1>. W OA *!
time/ppo/compute_rewards :'G1 OA *$
"
time/ppo/compute_advantages E;.&, E 0 OA *
time/ppo/optimize_stepm?sE,) 7_ OA *
time/ppo/calc_stats0T%>$ B+M OA *
time/ppo/totalN?F% 6 H OA *
env/reward_meanp=4$ B+M OA *
env/reward_std 9>#۹ " x= l^PA *
objective/kl3Bv' F ^PA *
objective/kl_coefP>W' F ^PA *
objective/entropy@6 / m] P ^PA *"
ppo/mean_non_score_rewardC9'% 6 ^PA *
ppo/mean_scores,q<>j4$ B+M ^PA *
ppo/std_scoresΪ>B Щ- ') 7_ "^PA *
ppo/policy/approxkl=-) 7_ "^PA *
ppo/policy/policykl *M) 7_ "^PA *
ppo/policy/clipfrac33<] F0 _ "^PA *#
!
ppo/policy/advantages_meanff*4!& sO #^PA *
ppo/returns/mean
a&% 6 #^PA *
ppo/returns/varIA,z_# wC ##^PA *
ppo/val/vpredTѾ# wC 3#^PA *
ppo/val/errorw_AB7z& sO #^PA *
ppo/val/clipfrac ?(c" x= '^PA *
ppo/val/mean p7! { '^PA *
ppo/val/varr@0+ K '^PA *
ppo/val/var_explainedpܽW.' F '^PA *
ppo/learning_rateo:\y+ K '^PA *
time/ppo/forward_passp/>x >. W '^PA *!
time/ppo/compute_rewards X:;1
(^PA *$
"
time/ppo/compute_advantages E;r, E (^PA *
time/ppo/optimize_stephl?
U) 7_ *(^PA *
time/ppo/calc_statsp$>[)<$ B+M ;(^PA *
time/ppo/total@?% 6 (^PA *
env/reward_mean,q<>.+$ B+M (^PA *
env/reward_stdΪ> ܼ&" x= ˃QA!*
objective/kl7Bbb' F 0QA!*
objective/kl_coefP>Gh' F KQA!*
objective/entropy`NAVACy/ m] P `QA!*"
ppo/mean_non_score_reward "6% 6 rQA!*
ppo/mean_scores=RLc$ B+M QA!*
ppo/std_scoresN=- % 6 ׄQA!*
ppo/loss/policy ;<$ B+M QA!*
ppo/loss/value?
$ B+M ɊQA!*
ppo/loss/totalN>%y( pJ QA!*
ppo/policy/entropy>B) 7_ QA!*
ppo/policy/approxkl@=`) 7_ #QA!*
ppo/policy/policykl,=<) 7_ 3QA!*
ppo/policy/clipfrac<ِ0 _ FQA!*#
!
ppo/policy/advantages_meanzUݯ& sO WQA!*
ppo/returns/mean8HҔ% 6 hQA!*
ppo/returns/var`@d# wC xQA!*
ppo/val/vpred2 nfB,# wC QA!*
ppo/val/error&c@"`& sO QA!*
ppo/val/clipfrac>=;$" x= QA!*
ppo/val/meant ! { ٍQA!*
ppo/val/varm?XL8+ K QA!*
ppo/val/var_explainedjC>' F QA!*
ppo/learning_rateo:sT+ K QA!*
time/ppo/forward_pass 0> . W (QA!*!
time/ppo/compute_rewards :LY1 QA!*$
"
time/ppo/compute_advantages TE;*j}, E QA!*
time/ppo/optimize_stephl?? 0) 7_ CQA!*
time/ppo/calc_stats$>YHv$ B+M QA!*
time/ppo/totalt?Rϟ% 6 QA!*
env/reward_mean=$ B+M AQA!*
env/reward_stdN=v@" x= BORA"*
objective/klR(B"@' F ORA"*
objective/kl_coefP>4' F ORA"*
objective/entropy>/ m] P ORA"*"
ppo/mean_non_score_rewardܾ_1% 6 ORA"*
ppo/mean_scoresd=g$ B+M PRA"*
ppo/std_scores"}=]~>- (( pJ VRA"*
ppo/policy/entropyb.:~
) 7_ VRA"*
ppo/policy/approxklc:) 7_ VRA"*
ppo/policy/policyklU:h) 7_ VRA"*
ppo/policy/clipfrac:i0 _ VRA"*#
!
ppo/policy/advantages_mean43k3;& sO WRA"*
ppo/returns/meanMY! 8% 6 WRA"*
ppo/returns/vara@T # wC &WRA"*
ppo/val/vpred<:$# wC 6WRA"*
ppo/val/erroraG?& sO YRA"*
ppo/val/clipfrac?U" x= YRA"*
ppo/val/mean\jap}! { YRA"*
ppo/val/var?x7+ K ZRA"*
ppo/val/var_explained(6?&' F ZRA"*
ppo/learning_rateo:pI + K %ZRA"*
time/ppo/forward_passP1>m'A. W 6ZRA"*!
time/ppo/compute_rewards h:Ͽ1 GZRA"*$
"
time/ppo/compute_advantages J;$, E WZRA"*
time/ppo/optimize_stepl?$) 7_ hZRA"*
time/ppo/calc_statsp$>=7$ B+M ZRA"*
time/ppo/totalJʡ?p% 6 [RA"*
env/reward_meand=͚$ B+M l[RA"*
env/reward_std"}=&
" x= _5TA#*
objective/kl\x(B' F 5TA#*
objective/kl_coef= Q>s77' F 5TA#*
objective/entropyX : ,/ m] P 5TA#*"
ppo/mean_non_score_reward&ܾM% 6 5TA#*
ppo/mean_scores,=$ B+M +5TA#*
ppo/std_scoresE= [- yp1$ B+M 5TA#*
ppo/loss/totalbp) 7_ 5TA#*
ppo/policy/approxkl-R) 7_ ~5TA#*
ppo/policy/policyklh4ҵ N) 7_ 5TA#*
ppo/policy/clipfrac t50 _ 5TA#*#
!
ppo/policy/advantages_mean926c& sO 5TA#*
ppo/returns/mean% 6 5TA#*
ppo/returns/var@# wC 5TA#*
ppo/val/vpredjڮGX# wC 5TA#*
ppo/val/error>& sO 5TA#*
ppo/val/clipfrac?" x= 25TA#*
ppo/val/mean:g ! { A5TA#*
ppo/val/vare.@E+ K :5TA#*
ppo/val/var_explaineduk?K' F k5TA#*
ppo/learning_rateo:ǣ+ K 5TA#*
time/ppo/forward_pass@>1>3&. W 5TA#*!
time/ppo/compute_rewards :O1 5TA#*$
"
time/ppo/compute_advantages J;G, E 5TA#*
time/ppo/optimize_stepul?r$) 7_ ,6TA#*
time/ppo/calc_stats@F%>RG$ B+M [6TA#*
time/ppo/totalꪡ?% 6 t6TA#*
env/reward_mean,= .|$ B+M 6TA#*
env/reward_stdE=
B|" x= gxUA$*
objective/kltx(BgR' F xUA$*
objective/kl_coef}+Q>A' F xUA$*
objective/entropy؎9h/ m] P xUA$*"
ppo/mean_non_score_rewardU>ܾ6% 6 yUA$*
ppo/mean_scoresc=8$ B+M $yUA$*
ppo/std_scoresr=կH- SWA& sO ~UA$*
ppo/val/clipfrac >" x= ʂUA$*
ppo/val/meanC|! { UA$*
ppo/val/varx@Ow+ K UA$*
ppo/val/var_explained0w?At~' F 'UA$*
ppo/learning_rateo:Ⱥ+ K UA$*
time/ppo/forward_passe1>P. W UA$*!
time/ppo/compute_rewards :^Ѐ1 ʅUA$*$
"
time/ppo/compute_advantages xE;(1K, E QUA$*
time/ppo/optimize_step l?ON) 7_ UA$*
time/ppo/calc_statsp$>jl3$ B+M UA$*
time/ppo/totalD?h?% 6 UA$*
env/reward_meanc=D39$ B+M UA$*
env/reward_stdr=Z9A" x= wVA%*
objective/klzx(BϜ' F VA%*
objective/kl_coefMQ> yp' F VA%*
objective/entropy9w/ m] P VA%*"
ppo/mean_non_score_rewardsbܾh3% 6 +VA%*
ppo/mean_scores=>$ B+M i% 6 VA%*
ppo/returns/var}@*S]K# wC VA%*
ppo/val/vpredʿR1# wC NVA%*
ppo/val/error` >#<& sO VA%*
ppo/val/clipfrac =L:" x= VA%*
ppo/val/meanпs(! { bVA%*
ppo/val/vari@ u+ K VA%*
ppo/val/var_explained[w?'T' F VA%*
ppo/learning_rateo:M_{+ K ~VA%*
time/ppo/forward_passТ.>Dt
. W VA%*!
time/ppo/compute_rewards @:[)81 FVA%*$
"
time/ppo/compute_advantages lB; , E VA%*
time/ppo/optimize_step^l?) 7_ VA%*
time/ppo/calc_stats$>uf@$ B+M YVA%*
time/ppo/total2?1q+% 6 VA%*
env/reward_mean=$ B+M VA%*
env/reward_std=: ro" x= ]XA&*
objective/klx(B_a' F ]XA&*
objective/kl_coef
pQ>' F ]XA&*
objective/entropy96,|/ m] P #]XA&*"
ppo/mean_non_score_rewardܾq6% 6 7]XA&*
ppo/mean_scoresЬ=oA$ B+M G]XA&*
ppo/std_scores }=Lw- + K &]XA&*
time/ppo/forward_passp0>j. W &]XA&*!
time/ppo/compute_rewards :̐2'1 &]XA&*$
"
time/ppo/compute_advantages F;4
, E ']XA&*
time/ppo/optimize_stepk?{,) 7_ ']XA&*
time/ppo/calc_stats$>\$ B+M {']XA&*
time/ppo/total4?p% 6 ']XA&*
env/reward_meanЬ=G~n$ B+M M(]XA&*
env/reward_std }=l9 " x= ,YA'*
objective/klx(B' F :-YA'*
objective/kl_coef^Q>mc' F _-YA'*
objective/entropybɜ9!2/ m] P t-YA'*"
ppo/mean_non_score_rewardڪܾn=% 6 -YA'*
ppo/mean_scores=??}$ B+M -YA'*
ppo/std_scores$5|=y<- ' F P5YA'*
ppo/learning_rateo:r+ K 5YA'*
time/ppo/forward_pass/> . W 6YA'*!
time/ppo/compute_rewards `:D1 ^6YA'*$
"
time/ppo/compute_advantages D;$J_, E 6YA'*
time/ppo/optimize_stepm?Yy) 7_ 7YA'*
time/ppo/calc_stats ">O{C$ B+M d7YA'*
time/ppo/total?DZ% 6 7YA'*
env/reward_mean=
$ B+M 8YA'*
env/reward_std$5|=K ]" x= LZA(*
objective/klx(B+p' F *MZA(*
objective/kl_coefQ>A' F LMZA(*
objective/entropy!97/ m] P _MZA(*"
ppo/mean_non_score_reward
ܾo(% 6 qMZA(*
ppo/mean_scoresW=(.x$ B+M MZA(*
ppo/std_scores(=>^5- . W UZA(*!
time/ppo/compute_rewards P:at1 HVZA(*$
"
time/ppo/compute_advantages ?; 9, E VZA(*
time/ppo/optimize_stepm?MvW) 7_ WZA(*
time/ppo/calc_stats#>Ai$ B+M \WZA(*
time/ppo/total?Xj% 6 WZA(*
env/reward_meanW=͠?L$ B+M XZA(*
env/reward_std(=