openai/train-procgen

Missing Code For Training on A Fixed Sequence of Levels

KarlXing opened this issue · 2 comments

According to the blog, you also have an experiment about the ablation with deterministic levels. If I'm correct, you set the use_sequential_levels label as True. Is it possible for you to release the code for that experiment? Or could you please give an example about environment setup? Thank you.

We haven't released code for that specific experiment, but I can walk you through what's necessary. You're correct that this requires setting the use_sequential_levels flag to True during training. You'll also need to set num_levels=1. Num levels is used to specify the number of possible seeds, and in this case we only want 1. Set start_level to any fixed value (this determines the particular seed you get). During testing, instead set num_levels=0 to remove the determinism. Also, remember to use the easy difficulty level distribution. To summarize:

training: {use_sequential_levels: True, num_levels: 1, start_level: YOUR_SEED, distribution_mode: "easy"}

testing: {use_sequential_levels: True, num_levels: 0, distribution_mode: "easy"}

@kcobbe may i know is above parameters work for procgen 0.10.3 i set according to above during training but still i get different rewards