Predicting human decisions with behavioral theories and machine learning

Ori Plonsky, Reut Apel, Eyal Ert, Moshe Tennenholtz, David Bourgin, Joshua Peterson, Daniel Reichman, Thomas L. Griffiths, Stuart Russell, Evan Carter, James F. Cavanagh, Ido Erev

Keywords: Choice under risk; Machine learning in behavioral science; Choice prediction competition; Behavioral economics; Model tournaments

TL;DR: BEAST-GB, a hybrid behavioral-ML model that predicts human choice under risk and uncertainty in STOA levels.

Abstract:

Accurately predicting human decision-making under risk and uncertainty is a long-standing challenge in behavioral science and AI. We introduce BEAST Gradient Boosting (BEAST-GB), a hybrid model integrating behavioral insights derived from a behavioral model, BEAST, as features in a machine learning algorithm. BEAST-GB won CPC18, an open choice prediction competition, and outperforms deep learning models on large datasets. It demonstrates strong predictive accuracy and generalization across experimental contexts, highlighting the value of integrating domain-specific behavioral theories with machine learning to enhance prediction of human choices.

Submission Number: 27