Brian Hlubocky and Eyal Amir
Computer text-based adventure games are virtual worlds in which humans or artificial agents act towards a specified goal through a text-based interface. In this paper we describe progress towards an agent that can interact with a game world in a human-like fashion. Precisely, we present the first accurate knowledge-gathering software agent that can track the state of the world in a textbased adventure game. This is nontrivial because such games are characterized by large, partially observable domains that include many objects, actions, and relationship between them. To test our agent, we developed a text-based adventure game world built by augmenting the LambdaMOO code base. We translated this game world into First-Order Logic, and used a logical filtering algorithm to allow the agent to track itself and the world efficiently. We expect that the development of software agents that act intelligently in such games will give rise to new types of games and will contribute to research on human-level artificial intelligence.