Heuristics for meta-planning from a normative model of information search

Planning, the process of evaluating the future consequences of actions, is typically formalized as search over a decision tree. More search means higher expected rewards, but tree search is computationally expensive. Most approaches designed to mitigate the costs associated with tree search have been driven by researcher-specified heuristics, and only recently have normative solutions been applied to the domain of planning. An open question is how people approximate the values associated with potential plans while they are planning. In this work, we propose to abstract planning as an information search problem to produce heuristics for meta-planning, or to determine which action to plan for. Specifically, we model a metacognitive process where evaluating candidate actions is viewed as gaining noisy measurements of the value of each action. This statistical estimate is then combined with prior experience in a Bayesian manner to decide whether and in which direction to continue sampling. This Bayesian meta-planner makes intuitive predictions across a range of parameters and acts as a more valuable, informed method for guiding search when compared to best-first and breadth-first search. Additionally, the Bayesian meta-planner qualitatively accounts for response time trends in a complex planning task. Thus, we provide a principled framework for information search that directs simulations towards the most promising actions, deriving heuristics that generalize to people's behavior while planning.