search

amidos2006 / gym-pcgrl

A package for "Procedural Content Generation via Reinforcement Learning" OpenAI Gym interface.
MIT License
113 stars 27 forks source link

Wide model #6

Closed smearle closed 4 years ago

smearle commented 4 years ago

Add the FullyConv2 NN used in our experiments.

amidos2006 commented 4 years ago

Why do you have wrapper called MaxStep?

I don't think we should have that so we don't give idea that we are limiting anything from outside the framework since the framework is limited internally by change percentage and max iterations which is just bigger value

smearle commented 4 years ago

I use MaxStep for inference, when I want to watch the agent play for a particular number of iterations, rather than limiting change % (which, as we've seen, can result in very different number of total iterations for different agents). I'm not using it for training or evaluation.

I find it helpful for qualitative evaluation (and I use it to make gifs). But no problem, I can keep it on a separate branch if preferred.

amidos2006 commented 4 years ago

Mmmm... But during inference, you can just make it a for loop instead of while done? the reason I am trying not to add new wrappers because it is just more confusing to understand the source code. I might just clean the wrappers and remove all unused wrappers so the repo is clean and easy to understand :D as we have bootstrap and tons of other stuff that we don't use anymore.

amidos2006 commented 4 years ago

The less stuff in the source code and less global functionality the easier the code to understand :) the more we have functions that do magics the more confusing it gets

smearle commented 4 years ago

Yes I agree! Let's trash all the excess code from the master. And thank you for pointing out that there's probably a much cleaner way to do MaxStep (lol).