Data augmentation for learning to play in text-based games

Abstract

Text-based game is an instance of partially observable environment where the observation and action are in the form of natural language. Generalizing in text-based games serves as a useful stepping-stone towards reinforcement learning (RL) agent with generic linguistic ability. Prior works on generalization in RL often applied data augmentation techniques, but none of them focused on text-based games. We propose a novel data augmentation technique for text-based games, Transition-Matching Permutation, where we identify phrase permutations that match as many transitions in the trajectory data. Applying this technique resulted in the state-of-the-art performance in a procedurally generated TextWorld's Cooking Game benchmark.