Postdoctoral Fellowship: SPRF: AI-based Models of Human Decision Making, Learning, and Innovation in Novel Systems of Rules and Reward

Collins, Katherine Maeve, Cambridge

Investigators

Abstract

Under the sponsorship of Dr. Joshua Tenenbaum at Massachusetts Institute of Technology, this postdoctoral fellowship award supports an early career scientist studying cognition and human behavior in novel systems of rules and reward. Systems of rules and reward govern many aspects of human behavior, from our jobs to markets to laws, institutions, and contracts, to sports, geopolitics, and mathematics. How is it that we can flexibly navigate a wide range of problems and social institutions that we have never encountered before -- and not just solve problems, but create new ones? Games provide a microcosm in which to study many aspects of cognition and human behavior in systems of rules and reward – and can be used to evaluate and build AI systems that are human-compatible. Much work around games in economics and AI focuses on settings where the gameplayer is an expert, perhaps having played many repeats of the same game; less work has studied how we play in, reason about, and even create new games. How do we flexibly make decisions in games we have never interacted with before and decide which games are worth thinking about? What underlies the earliest stages of learning -- after just a few rounds of play -- in a new game? And how do we decide which rules to bend, break, or create to make a game (or problem) more engaging? We posit that this involves abstract goal-directed evaluation functions and depth-limited stochastic search. We develop and empirically validate our theory through a series of large-scale behavioral studies, including synchronous human-human gameplay studies, over a wide range of novel games, including both competitive and cooperative settings. This project will computationally assess the hypothesis that much of the flexibility and productivity by which we think about games -- and multi-agent interaction in systems of rules and reward more broadly -- is powered by an “intuitive theory of games.” We will develop and empirically validate this theory through a series of large-scale behavioral studies, including synchronous human-human gameplay studies, over a wide range of novel strategy games, including both competitive and cooperative settings – and will develop models of human behavior, decision-making, and innovation using new AI-based computational methods. This proposal comprises three objectives: (Objective 1) assess how people decide what actions to take the first time they ever encounter a new strategy game; (Objective 2) model what happens in the early stages of learning as people shape strategies in new games; and (Objective 3) characterize how people innovate in new systems (creating and changing rules on-the-fly). This work will feed back to the development of more flexible human-like AI and more human-compatible interactive AI. We will also lay the foundations of a more cognitively grounded game theory which can support the design of more realistic simulations having implications for the science of national security and translational research in education and addiction. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.