Questions about Reinforcement Learning

I would greatly appreciate it if someone could answer my questions on the Reinforcement Learning examples. I would particularly like to know why the examples don't converge. Thank you.

Averill M. LAW

Answer 1

[Emmanouil Tzorakoleftherakis]

Hi Averill,

Can you please let me know which examples do not converge? I will share this information with the development team.

In the meantime, please have a look at this post. Every release comes with new features and optimizations under the hood, so the training plots may differ. The examples should still converge though (again please let me know if you find any examples that don't converge).

Answer 2

[Averill Law]

Thank you for your response. I asked several questions around May 9th. First of all, I ran the "Rocket Lander" example for the recommended 20,000 episodes and hyperparameters, and it was still having violent crashes. I have a Dell mobile work station (CPU) with a fast processor (2nd fastest available two years ago) and the example literally ran for 50 hours, rather than a few as stated in the literature.

I also could not get the "Stochastic Waterfall Grid World" example to converge, despite running it for thousands of episodes and many choices of the hyperparameters.

Averill M. LAW