Crashing beyond 6m iterations

[Coraku]

Hey!
Firstly, I really want to thank you for the great work you did, with the training paradigm successfully training on my 2017 Lenovo in about 1.2 minutes. All actors I trained performed positively, and I even found that your RL actors outperform the standard PID actor during wind, as the RL actor survived a wind experiment while the PID failed 5/5 times.

Now, during testing and fiddling with the iteration counter, I stumbled upon a problem with training longer than 6m iterations. I wanted to test the model for overfitting / diminishing returns, and had the idea to train an actor for an absurd amount of time. Sadly, the training script (I did the 3. approach of downloading the full repo) fully crashed (dumped core) when putting the training iteration variable to over 6 million.

I just wanted to share that, ask if this is a known issue or if you maybe know where this is coming from.
I'm very thankful for the project and based my Bachelors' thesis on it. Everything else worked great!

[jonas-eschmann]

Hey @Coraku

thank you for reaching out! That symptom is usually because of too high static memory / stack usage (it is crashing right away, not only after 6 million steps, right?).

Can you try increasing the EVALUATION_INTERVAL accordingly?

[Coraku]

Thanks for replying so quickly!
Yes, it crashes right away. Thanks for clarifying, I will try your fix and report back :)