deep-rl why is DDPG so unstable?

why is DDPG so unstable?

Open DontGiveUpEasily opened this issue 7 years ago • 2 comments

I can train a good agent, but the learning curve is quite noisy. why? is it an implementation issue or something intrinsic to DDPG?

Dec 18 '17 03:12 DontGiveUpEasily

@DontGiveUpEasily I trained it with 10k episode with the code, but the result is

Looks like it did not converge. Does the code work?

Jun 27 '18 23:06 GoingMyWay

@DontGiveUpEasily See my comment here: https://github.com/pemami4911/deep-rl/issues/2#issuecomment-400929047

Ideally, the OU noise needs to be decayed so that the actions don't have noise added after convergence

Jun 28 '18 07:06 Anjum48