Modeling Value-Based Decision-Making Policies Using Genetic Programming

An important way to develop models in psychology and cognitive science is to express them as computer programs. However, computational modeling is not an easy task. To address this issue, some have proposed using artificial-intelligence (AI) techniques, such as genetic programming (GP) to semiautomatically generate models. In this paper, we establish whether models used to generate data can be recovered when GP evolves models accounting for such data. As an example, we use an experiment from decision-making which addresses a central question in decision-making research, namely, to understand what strategy, or “policy,” agents adopt in order to make a choice. In decision-making, this often means understanding the policy that best explains the distribution of choices and/or reaction times of two-alternative forced-choice decisions. We generated data from three models using different psychologically plausible policies and then evaluated the ability and extent of GP to correctly identify the true generating model among the class of virtually infinite candidate models. Our results show that, regardless of the complexity of the policy, GP can correctly identify the true generating process. Given these results, we discuss implications for cognitive science research and computational scientific discovery as well as possible future applications.