Grant: Discovery Grants Program - Individual

Making decisions under varying degrees and types of uncertainty is a hallmark of human intelligence, and doing so in real-world settings at their full scope and complexity is one of the primary aspirations of artificial intelligence (AI). This uncertainty comes in varying forms: uncertainty about the outcome of future events, uncertainty about the true situation, and uncertainty about how other decision-makers will respond. There have been a number of recent advancements in applying AI decision-making systems to human-scale settings under various kinds of uncertainty. My research group has been pioneering new algorithms for coping with making decisions in settings where agents have asymmetric perspectives on the game. This work has led to solving human-scale poker games for the first time (2015), as well as programs able to beat professional poker players at ever more challenging poker variants (2017). We have also seen impressive advancements in situations where the rules of interaction are unknown, such as learning to play an game with no prior experience. The Arcade Learning Environment architected by my group (2013) is now one of the most widely used testbeds for evaluating learning agents for general competency. The testbed has stimulated breakthroughs by many research groups around the world, including Google DeepMind's celebrated human-level Atari algorithms combining deep learning and reinforcement learning (2015).These important advances in game theory and reinforcement learning are ripe to be translated into success outside of games and in real-world applications. However, there are still hurdles to overcome. The research in this proposal aims to address some of the shortcomings in current AI systems that prevent more widespread deployment. The goal is to move toward AI systems that can fully cope with real-world uncertainty in all its forms, that can effectively learn with less direct guidance, and that can better communicate their insight to human experts. Fundamental advances in these areas will be at the heart of the next generation of AI applications.