Experimental Evaluation of the Generalizability of the Soft Actor-Critic + Behavioral Cloning Algorithm

This repository is part of the Research Project CSE3000 in 2024 @ TU Delft. Other projects can be found here.

The research questions were:

1. Can SAC combined with BC effectively generalize to new tasks within a multi-task reinforcement learning environment?
2. What characteristics of the offline dataset are critical for the success or failure of SAC+BC in such settings?

Getting Started

1. Install Dependencies

pip install requirements -r requirements.txt

2. Open ./results.py

Follow Instructions in the main() Function in results.py

Environment:

The environment used is the MiniGrid Four-Room which contain 3 configurations, of which one is the training configuration the other are testing configuration for reachable and unreachable tasks. Every configuration will contain 40 tasks.

Datasets

The datasets have been created with an Optimal, Suboptimal and Mixed policy.

Optimal Policy - 100% success rate

Suboptimal Policy - 50% success rate (learned from a Deep Q-Network)

Mixed Optimal-Suboptimal Policy - 100% success rate