CSE
CSE
CSE CSE


Defense Event

Stable Profiles in Simulation-Based Games via Reinforcement Learning and Statistics

Mason Wright


 
Friday, November 30, 2018
10:00am - 12:00pm
Johnson Rooms A-C, Lurie Engineering Center

Add to Google Calendar

About the Event

In environments governed by the behavior of strategically interacting agents, game theory provides a way to predict outcomes in counterfactual scenarios, such as new market mechanisms or cybersecurity systems. Simulation-based games allow analysts to reason about settings that are too complex to model analytically with sufficient fidelity. But prior techniques for studying agent behavior in simulation-based games lack theoretical guarantees about the strategic stability of these behaviors.

In this dissertation, I propose a way to measure the likelihood an agent could find a beneficial strategy deviation from a proposed behavior, using a limited number of samples from a distribution over strategies, including a theoretically proven bound. This method employs a provably conservative confidence interval estimator, along with a multiple test correction, to provide its guarantee. I show that the method can reliably find provably stable strategy profiles in an auction game, and in a cybersecurity game from prior literature.

I also present a method for evaluating the stability of strategy profiles learned over a restricted set of strategies, where a strategy profile is an assignment of a strategy to each agent in a game. This method uses reinforcement learning to challenge the learned behavior as a test of its soundness. This study finds that a widely-used trading agent model, the zero-intelligence trader, can be reasonably strategically stable in continuous double auction games, but only if the strategies have their parameters calibrated for the particular game instance.

In addition, I present new applications of empirical game-theoretic analysis (EGTA) to a cybersecurity setting, involving defense against attacker intrusion into a computer system. This work uses iterated deep reinforcement learning to generate more strategically stable attacker and defender strategies, relative to those found in prior work. It also offers empirical insights into how iterated deep reinforcement learning approaches strategic equilibrium, over dozens of rounds.

Additional Information

Sponsor(s): Michael P. Wellman

Open to: Public