H H *JepA
brain.Event:2R.
,tensorboard.summary.writer.event_file_writerQ#" x= 9{pA*
objective/kl ؗ' F {pA*
objective/kl_coefL>uUX' F ({pA*
objective/entropy.B/ m] P @{pA*"
ppo/mean_non_score_reward ez% 6 U{pA*
ppo/mean_scoresͯ=$ B+M h{pA*
ppo/std_scores+_=/l[- ;U`7$ B+M Y{pA*
ppo/loss/value7A8JJ $ B+M ,{pA*
ppo/loss/total*?=( pJ {pA*
ppo/policy/entropyy@ <) 7_ {pA*
ppo/policy/approxkluHAc) 7_ ۤ{pA*
ppo/policy/policyklF@K) 7_ {pA*
ppo/policy/clipfracff>,E0 _ {pA*#
!
ppo/policy/advantages_mean43d7u& sO {pA*
ppo/returns/mean蛞?7% 6 +{pA*
ppo/returns/var>y# wC ;{pA*
ppo/val/vpred>`y# wC {pA*
ppo/val/error6AL& sO R{pA*
ppo/val/clipfrac>'w*" x= {pA*
ppo/val/meanN@ Zc! { {pA*
ppo/val/var4@91+ K D{pA*
ppo/val/var_explained'/' F ꩯ{pA*
ppo/learning_rateϸ: + K {pA*
time/ppo/forward_passP3>6d. W 4{pA*!
time/ppo/compute_rewards ԟ;Ky1 ֫{pA*$
"
time/ppo/compute_advantages x;AwU, E s{pA*
time/ppo/optimize_step ?JT) 7_ {pA*
time/ppo/calc_statsPl>)!$ B+M {pA*
time/ppo/total/?﹂% 6 E{pA*
env/reward_meanͯ=6EP$ B+M 㮯{pA*
env/reward_std+_=" x= sڜpA*
objective/kl{B_ ' F LtڜpA*
objective/kl_coef?L>[W' F tڜpA*
objective/entropy/B/ m] P tڜpA*"
ppo/mean_non_score_reward } % 6 tڜpA*
ppo/mean_scores =oI$ B+M uڜpA*
ppo/std_scores>u9- ɿO8( pJ G{ڜpA*
ppo/policy/entropy3?R;) 7_ ~ڜpA*
ppo/policy/approxklOAH6b) 7_ ڜpA*
ppo/policy/policykl|5X@P~) 7_ ڜpA*
ppo/policy/clipfrac >0 _ /ڜpA*#
!
ppo/policy/advantages_meanff3V& sO EڜpA*
ppo/returns/meanqx% 6 VڜpA*
ppo/returns/var Ag# wC օڜpA*
ppo/val/vpred4{:# wC ڜpA*
ppo/val/error\kBT& sO ڜpA*
ppo/val/clipfracT>u" x= ڜpA*
ppo/val/meanj ?{! { -ڜpA*
ppo/val/vars>*L;+ K ڜpA*
ppo/val/var_explained
z,' F @ڜpA*
ppo/learning_rateϸ:C(T+ K XڜpA*
time/ppo/forward_pass ~>X5G. W ڜpA*!
time/ppo/compute_rewards :91 ڜpA*$
"
time/ppo/compute_advantages m;Pp, E GڜpA*
time/ppo/optimize_stepR?S) 7_ ڜpA*
time/ppo/calc_stats0_>-
T$ B+M ڜpA*
time/ppo/total?J% 6 ڜpA*
env/reward_mean =I$ B+M ڜpA*
env/reward_std>C," x= PpA*
objective/klvRB)' F xPpA*
objective/kl_coefL>쌷P' F PpA*
objective/entropyjA/ m] P PpA*"
ppo/mean_non_score_reward=e % 6 $PpA*
ppo/mean_scorescD>}8$ B+M 7PpA*
ppo/std_scores >;- {0 _ PpA*#
!
ppo/policy/advantages_mean2327U& sO PpA*
ppo/returns/meanzz;% 6 -PpA*
ppo/returns/varPjA`H# wC PpA*
ppo/val/vpredoe# wC _PpA*
ppo/val/errorMHBV& sO PpA*
ppo/val/clipfrac43>]~" x= PpA*
ppo/val/meanT2! { !PpA*
ppo/val/vare:4+ K PpA*
ppo/val/var_explainedl'' F YPpA*
ppo/learning_rateϸ:Xc+ K PpA*
time/ppo/forward_pass0c>9. W PpA*!
time/ppo/compute_rewards :ei1 *PpA*$
"
time/ppo/compute_advantages V;w7, E PpA*
time/ppo/optimize_step&ӕ?
m) 7_ PPpA*
time/ppo/calc_statsPO>)$ B+M PpA*
time/ppo/total?{(r% 6 yPpA*
env/reward_meancD>ee<$ B+M
PpA*
env/reward_std >b2" x= 2.pA*
objective/kltEAb.' F U3.pA*
objective/kl_coefUL>ۻ' F x3.pA*
objective/entropy:N%!/ m] P 3.pA*"
ppo/mean_non_score_reward涝$% 6 3.pA*
ppo/mean_scoresoF>@$ B+M 3.pA*
ppo/std_scoresƔ;^`- . W v@.pA*!
time/ppo/compute_rewards :&1 XB.pA*$
"
time/ppo/compute_advantages U;dHVD, E B.pA*
time/ppo/optimize_step>ޕ?}d ) 7_ B.pA*
time/ppo/calc_stats0R>>$ B+M B.pA*
time/ppo/totalq? % 6 B.pA*
env/reward_meanoF>yy$ B+M B.pA*
env/reward_stdƔ;t" x=
pA*
objective/klEA @' F o
pA*
objective/kl_coefM>tr6' F
pA*
objective/entropyO9PB
=/ m] P
pA*"
ppo/mean_non_score_rewardНs% 6
pA*
ppo/mean_scores>q o$ B+M
pA*
ppo/std_scores9p:N- dV% 6 pA*
ppo/returns/var@
x# wC
pA*
ppo/val/vpred &# wC pA*
ppo/val/errorSA1& sO pA*
ppo/val/clipfracff&?^" x= pA*
ppo/val/mean f! { pA*
ppo/val/varxp:ڽe+ K pA*
ppo/val/var_explainedIK' F *pA*
ppo/learning_rateϸ:?H;+ K pA*
time/ppo/forward_pass ?e>)_. W $pA*!
time/ppo/compute_rewards 8:Jݰ1 BpA*$
"
time/ppo/compute_advantages Z;e, E pA*
time/ppo/optimize_step$?~ ) 7_ pA*
time/ppo/calc_statsP>s Q$ B+M pA*
time/ppo/totalN$?% 6 pA*
env/reward_mean>7$ B+M pA*
env/reward_std9p:o" x= A/pA*
objective/klEAg^' F /pA*
objective/kl_coef1M>U ' F /pA*
objective/entropy83:/ m] P /pA*"
ppo/mean_non_score_rewardZꝾ46% 6 /pA*
ppo/mean_scoresoF>eγv$ B+M
0pA*
ppo/std_scoresƔ;J1- $ B+M 66pA*
ppo/loss/total>\D!( pJ e6pA*
ppo/policy/entropyͼ:z) 7_ 6pA*
ppo/policy/approxkl\.7fE) 7_ 6pA*
ppo/policy/policykl/X 7(g) 7_ 6pA*
ppo/policy/clipfrac 0 _ 6pA*#
!
ppo/policy/advantages_meanff37 b& sO 6pA*
ppo/returns/meanjmL% 6 6pA*
ppo/returns/varM@<# wC 8pA*
ppo/val/vpredm[a# wC 9pA*
ppo/val/error5@`_& sO :9pA*
ppo/val/clipfrac43#?z:Yy" x= L9pA*
ppo/val/mean/r! { ]9pA*
ppo/val/varȑ:&\+ K n9pA*
ppo/val/var_explained%_' F ~9pA*
ppo/learning_rateϸ:9+ K 9pA*
time/ppo/forward_pass0^>{ k. W 9pA*!
time/ppo/compute_rewards :mX1 X:pA*$
"
time/ppo/compute_advantages h\;K , E :pA*
time/ppo/optimize_stepz?JY ) 7_ ;pA*
time/ppo/calc_stats@`K>N $ B+M |;pA*
time/ppo/total ?) % 6 ;pA*
env/reward_meanoF>@$ B+M * ' F % 6 [$ B+M >o( pJ ?. W Ak($ B+M 2$ B+M 6z' F -pA*
objective/entropy6:(%/ m] P -pA*"
ppo/mean_non_score_rewardV % 6 -pA*
ppo/mean_scores>3.o$ B+M -pA*
ppo/std_scores9p:3$^:- 9( pJ -pA*
ppo/policy/entropy3Cf:l
R) 7_
-pA*
ppo/policy/approxkli-]) 7_ H
-pA*
ppo/policy/policyklW6o) 7_ -pA*
ppo/policy/clipfrac PD0 _ 0-pA*#
!
ppo/policy/advantages_mean p2^\9& sO M-pA*
ppo/returns/meanz=7% 6 b-pA*
ppo/returns/var\?@ð[# wC r-pA*
ppo/val/vpred6ӿ +# wC -pA*
ppo/val/errorB@F@]& sO -pA*
ppo/val/clipfracff>G`" x= -pA*
ppo/val/meancX&! { -pA*
ppo/val/varM: + K -pA*
ppo/val/var_explainedck^w' F 0-pA*
ppo/learning_rateϸ:|l+ K \-pA*
time/ppo/forward_passP]>KfX. W t-pA*!
time/ppo/compute_rewards `: a1 -pA*$
"
time/ppo/compute_advantages <_;G7o, E 8-pA*
time/ppo/optimize_stepd=?]ۙ]) 7_ Q-pA*
time/ppo/calc_statsJ>̋$ B+M -pA*
time/ppo/totalV?ߐ% 6 ,-pA*
env/reward_mean>nY$ B+M -pA*
env/reward_std9p:C," x= "pA *
objective/klCA*' F "pA *
objective/kl_coefmM>g8U' F 1 "pA *
objective/entropyN:&{/ m] P F "pA *"
ppo/mean_non_score_reward
7% 6 Y "pA *
ppo/mean_scores>%$ B+M j "pA *
ppo/std_scores9p::&T- 3t x( pJ ('"pA *
ppo/policy/entropyϕ8" x= *"pA *
ppo/val/meanr! { -"pA *
ppo/val/var3;;'/+ K J-"pA *
ppo/val/var_explainedK=%Z' F c-"pA *
ppo/learning_rateϸ:+ K v-"pA *
time/ppo/forward_pass\>a6. W -"pA *!
time/ppo/compute_rewards :ڃ1 -"pA *$
"
time/ppo/compute_advantages Q;L7, E P0"pA *
time/ppo/optimize_stepJL?) 7_ }0"pA *
time/ppo/calc_statspI>r+ $ B+M 0"pA *
time/ppo/total6?}I% 6 0"pA *
env/reward_mean>-D' F j qA
*
objective/entropy
:%o/ m] P j qA
*"
ppo/mean_non_score_rewardP|% 6 j qA
*
ppo/mean_scores>l$ B+M k qA
*
ppo/std_scores9p: #- \D" x= Bw qA
*
ppo/val/mean ! { Rw qA
*
ppo/val/var;e+ K dw qA
*
ppo/val/var_explainedXv=" U' F uw qA
*
ppo/learning_rateϸ:'+ K w qA
*
time/ppo/forward_pass _>. W Tx qA
*!
time/ppo/compute_rewards p:DG1 x qA
*$
"
time/ppo/compute_advantages a;S, E y qA
*
time/ppo/optimize_step8Q? ) 7_ Yy qA
*
time/ppo/calc_stats*H>r6T
$ B+M y qA
*
time/ppo/total?P% 6 z qA
*
env/reward_mean>6 ON$ B+M Yz qA
*
env/reward_std9p:_K" x= qA *
objective/kl-CAR[' F _qA *
objective/kl_coefM>'' F qA *
objective/entropyY:q/ m] P qA *"
ppo/mean_non_score_rewardj#H% 6 qA *
ppo/mean_scoresoF>baZ$ B+M qA *
ppo/std_scoresƔ;Q&- " x= qA *
ppo/val/meanlP! { @qA *
ppo/val/var6Xl' F %qA *
ppo/learning_rateϸ:k+ K qA *
time/ppo/forward_pass@ ^>ߡ(. W qA *!
time/ppo/compute_rewards `:n1 JqA *$
"
time/ppo/compute_advantages T;3I!, E qA *
time/ppo/optimize_step|?) 7_ qA *
time/ppo/calc_statspJ> "Φ$ B+M dqA *
time/ppo/totalz?3% 6 qA *
env/reward_meanoF>xf$ B+M !qA *
env/reward_stdƔ;O%" x= %b qA *
objective/klBA݄^~' F b qA *
objective/kl_coefM>ǖO$ B+M b qA *
ppo/std_scoresƔ;7- 6.wC
$ B+M e qA *
ppo/loss/value2r?&'$ B+M h qA *
ppo/loss/total=&f( pJ h qA *
ppo/policy/entropy :Xt) 7_ h qA *
ppo/policy/approxkl->}) 7_ m qA *
ppo/policy/policykl6nb) 7_ Pm qA *
ppo/policy/clipfrac u.Y0 _ m qA *#
!
ppo/policy/advantages_mean1[ & sO m qA *
ppo/returns/meanV/% 6 m qA *
ppo/returns/var"@ # wC m qA *
ppo/val/vpred*E~lf,# wC m qA *
ppo/val/error ,?& sO n qA *
ppo/val/clipfrac=T2Z" x= n qA *
ppo/val/mean U~{! { 'n qA *
ppo/val/var=ȶ+ K q qA *
ppo/val/var_explainedn>k' F %r qA *
ppo/learning_rateϸ:ݶ
+ K >r qA *
time/ppo/forward_passB]>8<. W Rr qA *!
time/ppo/compute_rewards :y61 dr qA *$
"
time/ppo/compute_advantages a;I.?, E vr qA *
time/ppo/optimize_step~?@Z3) 7_ r qA *
time/ppo/calc_statsI>N$ B+M r qA *
time/ppo/total(?% 6 r qA *
env/reward_meanoF>vS$ B+M s qA *
env/reward_stdƔ;/N' F ,qA
*
objective/entropy:@gU/ m] P ,qA
*"
ppo/mean_non_score_reward%M@% 6 ,qA
*
ppo/mean_scoresoF>$ B+M ,qA
*
ppo/std_scoresƔ; [Y- " x= ,qA
*
ppo/val/meanf93~8! { },qA
*
ppo/val/varZ=ry\+ K q,qA
*
ppo/val/var_explained:>_' F ,qA
*
ppo/learning_rateϸ:mu+ K ,qA
*
time/ppo/forward_pass:`>Ek. W ,qA
*!
time/ppo/compute_rewards :F1 ,qA
*$
"
time/ppo/compute_advantages 0^; , E ,qA
*
time/ppo/optimize_step??C) 7_ `,qA
*
time/ppo/calc_statsI>~ $ B+M N,qA
*
time/ppo/total )?5 % 6 ~,qA
*
env/reward_meanoF>jD$ B+M ,qA
*
env/reward_stdƔ;
z" x= `8qA*
objective/kl>AA`' F `8qA*
objective/kl_coef?N>' F `8qA*
objective/entropy';/ m] P `8qA*"
ppo/mean_non_score_rewardSy8% 6 `8qA*
ppo/mean_scoresoF>/ʑ$ B+M `8qA*
ppo/std_scoresƔ;=- . W l8qA*!
time/ppo/compute_rewards :sh1 `l8qA*$
"
time/ppo/compute_advantages Z;, E l8qA*
time/ppo/optimize_step?͙f) 7_ m8qA*
time/ppo/calc_stats I>z$ B+M fm8qA*
time/ppo/totala? % 6 m8qA*
env/reward_meanoF>[M>$ B+M n8qA*
env/reward_stdƔ;'" x= .|DqA*
objective/klw
A§' F |DqA*
objective/kl_coef`N>*.' F |DqA*
objective/entropy> / m] P |DqA*"
ppo/mean_non_score_reward R% 6 |DqA*
ppo/mean_scores㯚>$r$ B+M |DqA*
ppo/std_scores'y=9q- $ B+M |DqA*
ppo/loss/value)>%M$ B+M }DqA*
ppo/loss/total6{=蹩( pJ I}DqA*
ppo/policy/entropy;;`+) 7_ c}DqA*
ppo/policy/approxkl>9{+֚. W
}DqA*!
time/ppo/compute_rewards 0:
:1
}DqA*$
"
time/ppo/compute_advantages DX;0.
, E 7
}DqA*
time/ppo/optimize_step.?#l') 7_ e
}DqA*
time/ppo/calc_statsJ>螘$ B+M O}DqA*
time/ppo/totalA?=% 6 |}DqA*
env/reward_mean㯚>+C$ B+M }DqA*
env/reward_std'y=?f