Using MDP similarity to increase generalization performance for Meta-RL agents

Meta-Reinforcement Learning (Meta-RL) addresses the limitations of standard reinforcement learning by enabling agents to quickly adapt to new, unseen tasks. Instead of learning a fixed policy, Meta-RL focuses on training agents to acquire adaptable learning strategies by practicing on a set of available tasks. Traditionally, tasks are sampled uniformly during training, but this may not lead to optimal learning efficiency.
This thesis will explore whether selecting tasks based on task similarity metrics, by estimating the distance between them, can improve the training process. The goal is to investigate whether more informed task selection can accelerate adaptation and lead to more effective Meta-RL.