Abstract: Developing agents capable of learning complex human-like behaviors is a key goal of artificial intelligence research. Progress towards this goal has exciting potential for applications in video games, from new tools that empower game developers to realize new creative visions, to enabling new kinds of immersive player experiences. In this talk I will share insights recently developed by my team and myself on how reinforcement learning, and other AI techniques can give rise to more human-like bot or NPC behaviors, focusing on the following components. First, I motivate the need for accurately evaluating human-likeness, propose a solution within a single-agent navigation task, and show that achieving highly skilled behavior is insufficient for human-likeness. Second, I demonstrate the first agent that passes our bar for human-likeness. Finally, I discuss how to go beyond single agent tasks and towards learning to collaborate, using structured models for predicting other agents’ behavior.

Bio: Katja is a Principal Researcher and lead of Game Intelligence at Microsoft Research Cambridge. Her research focuses on reinforcement learning, driven by current and future applications in video games, with the long-term goal of developing AI systems that learn to collaborate with people, to empower their users and help solve complex real-world problems. Her team is behind Project Malmo, a sophisticated AI experimentation platform built on top of Minecraft. She completed her PhD in Computer Science as part of the Information and Language Processing Systems group at the University of Amsterdam, supervised by Maarten de Rijke and Shimon Whiteson.