読者です 読者をやめる 読者になる 読者になる

Itsukaraの日記

最新IT技術を勉強・実践中。最近はDeep Learningに注力。

【DRL】Monezuma's Revengeの平均点が2000点に迫る

以前の記事を書いた後で、Montezuma's Revengeの平均点を上げようと色々ソースコードをいじっているため、以前に出た点数が出なくなっている可能性もあり、一応、以前の実行結果を再現させてみました。

今回、100M stepsまで実行し、平均点が2000点にかなり近づきました(学習曲線参照)。別途出力してる情報(ベスト平均点参照)では、78-79M steps近辺でのベスト平均得点は1520 - 1754でした。なお、平均点が一番高い78M steps近辺での学習データが残っていたので、githubにアップロードしました。ファイル名は「checkpoints.montezuma-c-avg-greedy-rar025-78M.tgz」です。

学習曲線

f:id:Itsukara:20161007134815p:plain

ベスト平均点(78.459M - 79.864M steps)

<Average Episode score (36 lines @ 78.459M - 79.864M STEP)>
@@@ Average Episode score = 1683.000000, s= 78459385,th=2
@@@ Average Episode score = 1570.000000, s= 78474074,th=5
@@@ Average Episode score = 1600.000000, s= 78534232,th=3
@@@ Average Episode score = 1528.000000, s= 78534463,th=1
@@@ Average Episode score = 1561.000000, s= 78602286,th=0
@@@ Average Episode score = 1730.000000, s= 78649996,th=7
@@@ Average Episode score = 1596.000000, s= 78763734,th=4
@@@ Average Episode score = 1539.000000, s= 78787510,th=6
@@@ Average Episode score = 1585.000000, s= 78790811,th=5
@@@ Average Episode score = 1723.000000, s= 78799980,th=2
@@@ Average Episode score = 1532.000000, s= 78864929,th=1
@@@ Average Episode score = 1561.000000, s= 78889484,th=3
@@@ Average Episode score = 1541.000000, s= 78962613,th=0
@@@ Average Episode score = 1754.000000, s= 78976248,th=7
@@@ Average Episode score = 1621.000000, s= 78996270,th=4
@@@ Average Episode score = 1619.000000, s= 79100351,th=5
@@@ Average Episode score = 1664.000000, s= 79115006,th=2
@@@ Average Episode score = 1529.000000, s= 79128643,th=6
@@@ Average Episode score = 1552.000000, s= 79202649,th=3
@@@ Average Episode score = 1573.000000, s= 79225666,th=1
@@@ Average Episode score = 1604.000000, s= 79308819,th=4
@@@ Average Episode score = 1734.000000, s= 79311025,th=7
@@@ Average Episode score = 1561.000000, s= 79323126,th=0
@@@ Average Episode score = 1680.000000, s= 79414686,th=5
@@@ Average Episode score = 1703.000000, s= 79427641,th=2
@@@ Average Episode score = 1600.000000, s= 79437383,th=6
@@@ Average Episode score = 1569.000000, s= 79558727,th=1
@@@ Average Episode score = 1592.000000, s= 79562370,th=3
@@@ Average Episode score = 1562.000000, s= 79625564,th=4
@@@ Average Episode score = 1520.000000, s= 79657699,th=0
@@@ Average Episode score = 1636.000000, s= 79668274,th=7
@@@ Average Episode score = 1678.000000, s= 79721917,th=5
@@@ Average Episode score = 1553.000000, s= 79739129,th=6
@@@ Average Episode score = 1668.000000, s= 79739390,th=2
@@@ Average Episode score = 1520.000000, s= 79784754,th=1
@@@ Average Episode score = 1546.000000, s= 79864012,th=3