Domain Adaptation for Deep Reinforcement Learning in Visually Distinct Games

Dino S. Ratcliffe, Luca Citi, Sam Devlin, Udo Kruschwitz

Abstract:

Many deep reinforcement learning approaches use graphical state representations,
this means visually distinct games that share the same underlying structure cannot
effectively share knowledge. This paper outlines a new approach for learning
underlying game state embeddings irrespective of the visual rendering of the game
state. We utilise approaches from multi-task learning and domain adaption in
order to place visually distinct game states on a shared embedding manifold. We
present our results in the context of deep reinforcement learning agents.

TL;DR: An approach to learning a shared embedding space between visually distinct games.

Keywords: Deep Reinforcement Learning, Domain Adaptation, Adversarial Networks