This paper introduces a novel architecture for reinforcement learning withdeep neural networks designed to handle state and action spaces characterizedby natural language, as found in text-based games. Termed a deep reinforcementrelevance network (DRRN), the architecture represents action and state spaceswith separate embedding vectors, which are combined with an interactionfunction to approximate the Q-function in reinforcement learning. We evaluatethe DRRN on two popular text games, showing superior performance over otherdeep Q-learning architectures. Experiments with paraphrased action descriptionsshow that the model is extracting meaning rather than simply memorizing stringsof text.
展开▼
