How could I realize from the graph of "reinforcement learning episode manager" in DDPG or rlTD3Agent that the network is learning well?

hi every body

im training a TG3Agent . when i see the "reinforcement learning episode manager" i dont realize that my network is learning well or not?

how can i found that? What are the signs of good learning from the chart?

my chart is below

Answer 1

[Emmanouil Tzorakoleftherakis]

You should be seeing the average reward curve going upwards (not monotonically) which is not the case here.

[ali kadkhodaei]

whats your advice for well learning ? increasing layers or nurons ?

using lstm layers instead of feauture layers?/

changing learning rate or gradient thereshhold?

or another things?

[Emmanouil Tzorakoleftherakis]

Using LSTM layers instead of feature layers won't improve your performance unless the problem you are working on is a better fit for lstm policies. Usually improving the reward, exploration, and neural network architecture (layers/neurons etc) helps. But you also need to make sure the problem is set up correctly, i.e. you provide the appropriate observations so that the agent has enough information to learn.