using common sense for decision making in an adventure game（使用常识进行决策的冒险游戏）.pdfVIP

Using Common Sense for Decision Making in an Adventure Game Deepak Ramachandran Computer Science Dept. University of Illinois at Urbana-Champaign Urbana, IL 61801 Abstract reward function to effect faster convergence). Our experi- ments show that using the correct knowledge to do reward In this paper we apply common sense knowledge to the task shaping leads to substantial improvements in the learning of decision making in an adventure game. We show how rate. knowledge about the underlying domain, induced from a Common sense is the central problem of Artificial Intel- common sense knowledge base, can be used to do reward shaping of the environment, enabling reinforcement learning ligence but the question of its practical use is poorly under- agents to learn optimal behaviours faster. The lessons learnt stood. Large Knowledge Bases such as CYC and Concept- from this experiment could be used as a first step to finding Net have been constructed in recent years, yet their applica- robust applications of common sense knowledge. tions to real-world problems are notoriously brittle. There is always the danger that using a particular axiom from a KB in a certain context could lea