Where communities thrive

Join over 1.5M+ people
Join over 100K+ communities
Free without limits
Create your own community

Explore more communities

chainer/rl

People

Repo info

Activity

Shunta Saito

@mitmul

-sameroom portal

Sameroom

@sameroom-bot

<Sameroom> Your Portal URL is https://sameroom.io/DnTuzTqM -- you can send the URL to someone on a different team to share this room. Note: you can connect more than two teams this way.

<Sameroom> I've connected 1 new room #rl (chainer) on Slack. See map

Sameroom

@sameroom-bot

[peisuke, chainer] 質問です。actに2種以上のデータは渡せるでしょうか。例えば、ロボットの場合ならば、画像とセンサ情報のtupleなどです。

Sameroom

@sameroom-bot

[not, chainer] @peisuke ここは英語用なので日本語で質問したいならchainer-jpで聞いた方が良いですよ
Because here is for English, you had better ask at chainer-jp slack if you want to ask in Japanese.

[peisuke, chainer] tnx, I posted question in Japanese. My question is, i want to pass two or more state variables to the act function. For example, image and sensor data. Is it possible?

Sameroom

@sameroom-bot

[Yasuhiro Fujita, chainer] You mean the act method of ChainerRL agents?

Sameroom

@sameroom-bot

[Bernard Lee, chainer] Hello, just wondering if anyone here had any experience using chainerrl's ddqn agent?

Sameroom

@sameroom-bot

[Yasuhiro Fujita, chainer] you mean DoubleDQN?

Sameroom

@sameroom-bot

[Bernard Lee, chainer] Yes, the double dqn

[Bernard Lee, chainer] I'm having a CNN as my Q-function and I seem to always have the problem where the Q-values explodes to very high once experience replay starts and then it slowly decreases back down.

Sameroom

@sameroom-bot

[Yasuhiro Fujita, chainer] DQN families are not stable algorithms and need careful tuning. At least I recommend using the same settings as DeepMind's Nature paper: scaling of the input value into [0-1] and reward [-1,1] as well as using the exact same hyperparameter settings.

Sameroom

@sameroom-bot

[Bernard Lee, chainer] Thank you

[Bernard Lee, chainer] Will try that