A Study of Datum Search Patterns Using a Stochastic Game Framework

The paper considers the autonomous datum search problem, that is, a search for an evader, starting with some delay from a (possibly uncertain) location at which it has been sighted. The context is underwater surveillance. We treat the evader as an intelligent player and cast the datum problem as an autonomous search using the framework of partially observable stochastic two-player zero-sum games. A realistic model of sensing is adopted, while (uncertain) knowledge of evader’s location is represented by a dynamic probabilistic occupancy map, updated via Bayes rule. The payoff assigned to each searcher-evader pair of actions is defined as a reduction of entropy of the probabilistic occupancy map. This game was played repeatedly ‘in silico’, for the purpose of determining the search patterns and search statistics. The main contribution is a discovery of search patterns in (until present) unsolved settings using a realistic sensing model.