A Point-Based Approximate Algorithm for One-Sided Partially Observable Pursuit-Evasion Games

Pursuit-evasion games model many security problems where an evader is trying to escape a group of pursuing units. We consider a variant with partial observability and simultaneous moves of all units, and assume the worst-case setup, where the evader knows the location of pursuer's units, but the pursuer does not know the location of the evader. Recent work has shown that the solution of such games is compactly representable as a collection of finite-dimensional value functions. We extend this result and propose the first practical algorithm for approximating optimal policies in pursuit-evasion games with one-sided partial observability. Our approach extends the point-based updates that exist for POMDPs to one-sided partially observable stochastic games. The experimental evaluation on multiple graphs shows significant improvements over approximate algorithms that operate on finite game trees.