Investigating Human Learning and Decision-Making in Navigation of Unknown Environments

Humans often navigate in unknown and complex environments. As they gain experience, they can eventually determine near-optimal (e.g., minimum-time) paths between two locations from memory. The goal of this research is to understand the heuristics that humans use to solve path-planning problems in unknown environments. This paper presents a modeling and analysis framework to investigate and evaluate human learning and decisionmaking while learning to navigate unknown environments. This approach emphasizes the agent (a vehicle with a human driver on board) dynamics, which is not typical in navigation studies. The framework is based on subgoals that are defined as intrinsic patterns in interactions between agent dynamics and task environment. Subgoals represent nodes in a graph representation of the task space. The evaluation framework uses Dijkstra's algorithm to find minimum-time paths in the subgoal graph. To account for limited working memory in humans, the shortest-path search in the graph is terminated at a specified maximum depth. The cost beyond the maximum depth is approximated using learned cost-to-go values at subgoals. The graph framework is applied to evaluate human data from simulated guidance experiments in which subjects were asked to find minimum-time routes from pre-specified start to goal states, over multiple trials.