Explainable RL
solving atari games with neurosymbolic explainable plans
Overview
This project focuses on explainable reinforcement learning by developing object-centric representations and surrogate models to interpret the decision-making of Deep Q-Network (DQN) agents trained on ATARI games.
Key Contributions
- DQN Training & Object-Centric Representations
- Trained DQN agents on ATARI environments, Pong,Boxing.
- Contributed to a neuro-symbolic RL framework for object-centric learning.
- Surrogate Models for Explainability
- Distilled DQN behavior into decision trees and neural networks for exploring imitation implications.
- Leveraged object-centric data to enhance interpretability of RL agents.
- Guided Exploration for RL Agents
- Currently working on region proposal methods to improve exploration strategies in ATARI games.
- This work focuses on code based meta-controller for hard exploration games like PitFall and Montezuma’s revenge
Acknowledgment
Conducted at the AIML Lab, TU Darmstadt, under supervision of Quentin Delfosse and Prof. Kristian Kersting as a HiWi(Funded research assistant position)