Individual research project · 2025
Offline replay in tabular reinforcement learning
Tabular reinforcement learning with replay and perturbed experience in a small GridWorld.
Conceptual illustration; not an experimental result.
Research question
Could replaying slightly altered experiences help a reinforcement learning agent adapt when its environment changes? This study examines the idea with a deliberately small, inspectable setup.
Methodology
The final report describes tabular Q-learning in a 6 × 6 GridWorld. Experience replay and dreaming-style perturbations are compared with a baseline, examining acquisition and transfer behaviour.
Findings
The report describes slower initial acquisition, alongside limited indications of transfer benefits. Those observations are specific to this small experiment.
Limitations
This is a tabular reinforcement learning study. The evidence does not establish a general advantage for learned world models or large-scale agents. The report is the source for the implementation and findings presented here.
Related research
Sign language recognition