Closed esalehi1996 closed 1 year ago
Hi @esalehi1996 thank you for bringing this up. We are working on a PR to add this as a feature, #265 . You'll be able to initialize any minigrid environment with the max_steps
argument to set the total number of steps per episode.
How can I initialize a MiniGrid environment with a custom max_steps value?
My issue is that I believe that the default max_steps value for the Multi-room family of environments is maybe a bit low (120 for the 6 room environment).
I am running some experiments using a slightly modified recurrent Q-learning approach (similar to the r2d2 paper). I have been able to solve harder (I assume) environments such as the ObstructedMaze-2Dlhb and KeyCorridorS4R3 environments with my approach but my agent is unable to learn anything in MultiRoom-N6 simply because episodes end very quickly and there is no episode with non-zero reward in the replay buffer.
Any help is appreciated.
I am assessing the effectiveness of a model-augmented recurrent Q-learning approach versus a vanilla recurrent Q-learning approach (r2d2) and I testing my approach on all the environments in the MiniGrid family. (So far I've seen some big improvements especially in the ObstructedMaze and KeyCorridor family).