Using Walkthroughs to Improve Transfer Learning in Text-Based Gameplay

Abstract

Text-based games have the potential to facilitate the learning of environmentally-grounded language representations. A number of agents already exist that can play individual text-based games with some level of success, however none to date have demonstrated a sustained ability for their learned language representations to transfer over to improved gameplay in dissimilar games. This suggests that the agents are not learning general language representations that are useful outside of the context of individual games. To remedy this problem, we propose the use of walkthroughs as a training guide to help agents encounter a wider variety of high-leverage states and learn representations that transfer to other games. In particular, we both built a unique supervised model to translate natural language walkthroughs into explicit game com- mands and tested a novel strategy for incorporating walkthroughs into reinforcement learning gameplay. Ultimately, our results were mixed: our supervised model was mostly successful at outputting accurate game commands while our reinforcement learning agent did not see im- proved results as a result of incorporating walkthroughs into its training.