Transferable state & action embeddings in deep reinforcement learning

Abstract

This thesis delves into two primary areas: the impact of transfer learning on the performance of RL algorithms and the implementation of a transformers-based architecture for learning representations of both states and actions. The proposed architecture, named JASE-DQN (Joint Action and State Embeddings Deep Q-Network), introduces a novel approach by concurrently learning latent representations of state and action attributes, as well as their interdependencies. Transfer learning process is facilitated through the integration of pretrained models for state and action representations within a cross-attention transformer framework.