I Can't Think With All This Noise: Inferring Strategies Using Symbolic Regression

We use symbolic regression (implemented by a genetic program) to analyze the role of agent expectation formation in games. In the model, agents attempt to infer the strategies of opponents through regression and then best respond using this information. Though agents use deterministic strategies, behavior that resembles a mixed strategy emerges. When one agent uses more complicated strategy primitives (building blocks) significant performance advantages are realized. However, even small amounts of noise in the system can eliminate this advantage. By changing the design of the game we show that it is crucial to accurately infer past actions in order to realize performance advantages from complexity.