Representation Learning for Behavioral Analysis of Complex Competitive Decisions

Abstract

Many phenomena studied in marketing and economics are analyzed through the lens of non-cooperative games, but research on complex real-world competitive settings is rare due to methodological challenges and data limitations. To bridge this gap, I develop a neural network architecture that enables behavioral analysis of complex games by estimating a game's payoff structure (e.g., win probabilities between pairs of actions) while simultaneously mapping player actions to a lower-dimensional latent space with a linear structure that enables behavioral analysis. I apply my method to a unique dataset of over 11 million matches of a competitive video game (Pokemon VGC) with a large array of actions and complex strategic interactions. I find that players select actions that counterfactually would have performed better against recent opponents, demonstrating model-based reasoning. Still, players overweight simple heuristics relative to model-based reasoning to an extent that is similar to prior findings in lab settings. I find that noisy and biased decision-making leads to frequent selection of suboptimal actions, which corresponds to lower player engagement. This demonstrates the limits of player sophistication when making complex competitive decisions and suggests that platforms hosting competitions may benefit from interventions that enable players to improve their decision-making.

Keywords: machine learning, representation learning, decision-making, behavioral game theory

JEL Classification: C45, C55, C57, C72, C73, D90