RL for navigation

Overview

This is a repo for training reinforcement-learning agents is a gridworld, such that agents optimize for the shortest integrated Euclidean distance to rewards. Currently supported models:

• Tabular Q learning
• Tabular SARSA
• State-action Successor Representation
• Tile coding (with Q learning)
• Hierarchical state space (with Q learning or SARSA)
• Model-based tree search

Usage

• Install the repo from the directory ...\Euclidean_Gridworld_RL using pip install -e .
• Modify the config files as needed in experiments/suites/...
• Run an experiment, e.g. with the command python run.py --mode parallel --seeds 10 --config_path suites/a_star/config.yaml --config_changes config_changes.py
• Visualize results, e.g. with the command python post_process.py --plot_t --results_folder results\...