Proceedings of the Annual Meeting of the Cognitive Science Society

Searching natural environments, as for example, when foraging or looking for a landmark, combines reasoningunder uncertainty, planning and visual search. Existing paradigms for studying search in humans focus on step-by-step infor-mation sampling, without examining advance planning. We propose and evaluate a Bayesian model of how people search in anaturalistic maze-solving task. The model encodes environment exploration as a sequential process of acquiring informationmodelled by a Partially Observable Markov Decision Process (POMDP), which maximises the information gained. We showthat the search policy averaged across participants is optimal. Individual solutions, however, are highly variable and can beexplained by two heuristics: thinking and guessing. Self-report and inference, a Gaussian Mixture Model over inverse POMDP,consistently assign most subjects to one style or the other. By analysing individual participants’ decision times we show thatindividuals solve partial POMDPs and plan their search a limited number of steps in advance.