Why the Reinforcement Learning seems do not learn anything?

Is reward not converge to a certain value show that the RL agent does no learn anything?

The result shows that every training the agent does the different choices, it won't learn something good from the previous one.

Although the reward is good and has the good result, next training it won't keep at that good choices, it will try the other choice then get the bad result.

How can I deal with this problem?

Thank for helping.

Answer 1

[Tarunbir Gambhir]

If the agent is not taking good choices at later episodes, it is likely that the exploration epsilon factor is still high. You can try increasing the "agentOptions.EpsilonGreedyExploration.EpsilonDecay" parameter to encourage the agent to exploit the previously learned Q-values at later episodes.

You can refer this documentation page for more information on the importance of parameters for the epsilon-greedy exploration concept.