[3F5-ES-2] Machine learning: Conversation and emotions

[3F5-ES-2-05] Learning to Set Planning Goals for TextWorld by Deep Neural Semantic Parsing

〇Don Joven Agravante¹, Michiaki Tatsubori¹ (1. IBM Research Tokyo)

キーワード：neuro-symbolic AI, ML applications in game playing

Our task is learning to set the goals of a planner to solve text-based adventure games, specifically TextWorld, where each game has a different quest that is described in natural language. Translating the quest into a logical form consistent with the PDDL is a semantic parsing problem which we tackled by training a Transformer neural network in a supervised way. We then show how a game-playing agent can be made by using the neural network with a planner and some external knowledge. Our results show that this agent can solve the most complex class of the TextWorld game settings, including sparse rewards. This agent architecture bridges the gap between neural networks and classical planning in a novel way by grounding the Transformer output into the PDDL symbolic layer.