Committee: Ambuj K. Singh (chair), Francesco Bullo, Noah Friedkin.
Title: Online Learning in Group Decision Making
Abstract:
 

A perfect agent acting within a complex environment incorporates all relevant data and uncertainties, both tangible and intangible, when making decisions. In reality, even the most adept decision makers employ simplifications, heuristics, and partially rational choices due to noisy data, cognitive biases, and limited processing power. Further, interacting agents must also contend with the irrationalities introduced by other one another when acting.

In this talk, I will first explore how game theory and behavioral economics explain such deviation from rationality. For example, Prospect Theory develops a simple, yet powerful model for explaining the bounded rationality of human decision making. This will include a review of recent work on modeling humans as prospect theoretic agents to gain insights about how bounded rationality affects decision making. Such work covers diverse topics such as microgrid energy markets, cyber-physical security for drone deliveries, binary hypothesis testing, and encoding valuable information over a noisy channel.

Second, I will briefly cover decision making from an online learning perspective. Typically, a decision maker designs a policy to interact with a noisy, but rational, environment in order to optimize a performance function. Consequently, benchmarks of success in this field are computationally efficient policies which enjoy theoretically-guaranteed performance. I will discuss the potential and limitations of state-of-the-art methods in complex settings .

Finally, I will discuss the challenges and potential impact of fusing these approaches. We posit the primary challenge of a unified approach is reconciling the steady-state analysis of game theory with the data-driven model building required for online policy design in interactive environments. We propose that a successful fusion would allow decision makers to accurately assess their limitations and improve their decision making with data-driven models and theoretically-sound policies.