Text-based RL Agents with Commonsense Knowledge: New Challenges, Environments and Baselines

Text-based games have emerged as an important test-bed for Reinforcement Learning (RL) research, requiring RL agents to combine grounded language understanding with sequential decision making. In this paper, we examine the problem of infusing commonsense knowledge. Such knowledge would allow efficiently act in world by pruning out implausible actions, and perform look-ahead planning determine how current actions might affect future states. We design a new text-based gaming environment called TextWorld Commonsense (TWC) training evaluating specific kind about objects, their attributes, affordances. also introduce several baseline which track context dynamically retrieve relevant from ConceptNet. show that incorporate TWC better, while acting more efficiently. conduct user-studies estimate human performance on there is ample room improvement.