openai / baselines

OpenAI Baselines: high-quality implementations of reinforcement learning algorithms
MIT License
15.83k stars 4.88k forks source link

Question on VecNormalize wrapper "training" parameter.. #1227

Open Hydraxx10 opened 11 months ago

Hydraxx10 commented 11 months ago

Hello, I have an A2C model that i trained using VecNormalize wrapper and when evaluating using a another environment, the Boolean "training" parameter gives different results going from training == False to training == True, I saw the documentation saying that this parameter allows "Whether to update or not the moving average" but, I still don't know if keeping the training as True is consider cheating myself. So, I need help.

image