Partial Observable Grid World Experiments

This repository contains experiments for partially observable grid world environments in GridVerse with Stable Baselines, and RLlib.

Key Feature

Custom networks for gridverse dict observations. This network at a high level is a CNN followed by an optional RNN. To make this work, custom encoder where implmeneted in stabelines and custom catlog, connectors and encoders in rllib. Hydra is used for config management.

Directory structure

• config (top level) is for configuring gridverse environment (rewards, states, observations and such) and for configuring the algorithm and hyperparameters.
• ray_custom/experiment_manager.py is the driver code for running PPO( easily configurable to other algorithms from rllib) in gridverse with rllib
• stablebaselines_exp/stablebaselines_exp/stablebaselinestrain.py is the driver code for training PPO( easily configurable to other algorithms from stablebaseline) in grifverse with stablebaseline
• gridverse_torch_featureextractors are pytorch networks that are used as building pieces in both rllib and stablebaselines.
• gridverse_utils contains some utility functions for creating the gridverse environment as a gym environment.