Deep Reinforcement Learning with a Natural Language Action Space
This addresses the challenge of handling natural language actions in reinforcement learning for text-based games, representing an incremental improvement with a novel method for a known bottleneck.
The paper tackles the problem of reinforcement learning in text-based games with natural language action spaces by introducing a deep reinforcement relevance network (DRRN), which outperforms other deep Q-learning architectures on two popular games and demonstrates the ability to extract meaning from paraphrased actions.
This paper introduces a novel architecture for reinforcement learning with deep neural networks designed to handle state and action spaces characterized by natural language, as found in text-based games. Termed a deep reinforcement relevance network (DRRN), the architecture represents action and state spaces with separate embedding vectors, which are combined with an interaction function to approximate the Q-function in reinforcement learning. We evaluate the DRRN on two popular text games, showing superior performance over other deep Q-learning architectures. Experiments with paraphrased action descriptions show that the model is extracting meaning rather than simply memorizing strings of text.