Solving Physical Traveling Salesman Problems with Policy Adaptation

The Physical Traveling Salesman Problem (PTSP) is a current research problem which adds a model of velocity to the classic TSP. In this paper we propose algorithms for solving the PTSP which avoid the fragmented allocation of memory and precompute cell-precise single-source shortest paths for each waypoint by using an engineered implementation of Dijkstra's algorithm. To determine an initial tour, we solve ordinary and general TSPs. For moderately sized problems, we apply an optimal depth-first branch-and-bound TSP solver which warrants constant-time per search tree node. For larger problems, we apply randomized search with policy adaptation to learn from good tours. We evaluate our solution with a series of benchmark experiments and compare the results to the winner of the PTSP competition at CIG 2013. In comparison, our approach shows similar results but also provides a graph search with optimal time performance.