Behaviour Characterization via Policy Supervectors in Multi-Task Reinforcement Learning

This thesis proposes a novel Multi-Task Reinforcement Learning (MTRL) framework that facilitates inter-environment policy transfer by leveraging state-based behavioural characterizations. Our method encodes policies through supervector representations derived from Gaussian Mixture Models (GMMs) trained on visited states. By embedding policies into a shared latent space, we enable similarity-based policy selection, guiding learning in new tasks based on previously acquired knowledge. This approach is evaluated across environments from the Highway-Env suite, aiming to demonstrate its effectiveness in accelerating adaptation, improving generalization, and identifying transferable skills between source and target tasks.