I was recently considering training an agent that perform a task by reinforcement learning. Both the state and actions are continuous, but could be discretized if needed. The problem is that in my case the state observations and reward will be quite noisy, so given the same state and action, the next state and received reward will be different on each run, and the noise cannot be described by canonical probability distributions.

Up to now I have tried deep Q-network, stochastic policy gradient and deep deterministic policy gradient. While I could successfully implemented these algorithms in the CartPole game, they all failed to learn my particular task.

I hope to know are there any reinforcement learning methods that can deal with noisy state observations?

  • $\begingroup$ I failed to understand your question. If your rewards are not correlated to the observations by any means, then how would the agent learn? $\endgroup$
    – Yahya
    Commented Aug 1, 2021 at 21:27
  • $\begingroup$ @Yahya It is not totally uncorrelated, the state observation and rewards are just corrupted by some noise. As a example, consider letting a human to estimate the distance between him/her and a object. Across different trials the person may report different values even if the actual distance has not changed, but the mean reported distance should be somehow correlated with the actual distance. $\endgroup$
    – Cloudy
    Commented Aug 2, 2021 at 5:25


You must log in to answer this question.

Browse other questions tagged .