Modeling Goal Selection with Program Synthesis

In reinforcement learning, it can be difficult to select goals among many possible states. We define a framework for understanding optimal goal selection and its computational cost. We then propose program induction as a method for defining human-like priors that make informed goal selection easier. By generating programs that map to a state space and reward function, we efficiently approximate an optimal goal selecting agent. We highlight applications of this work to sequential goal selection and modeling of human behavior.