Expected reward blows up while training (DDPG agent, reinforcement learning)

I am training a DDPG network and after training for around 5000 iterations, the model seems doesnot seem to converge while the expected reward keeps on increasing exponentially. What can be a possible reason and how to solve the issue.

Answer 1

[Emmanouil Tzorakoleftherakis]

Hello,

This answer may be helpful.

I would make sure your reward signal outputs values that make sense, and also possibly simplify the critic network.

[Sayak Mukherjee]

Thanks for your answer

What does simplifying critic network mean? Does that mean use less nodes and hidden layers?

[Emmanouil Tzorakoleftherakis]

That's right