Finding the Needle in the Haystack with Heuristically Guided Swarm Tree Search

In this paper we consider the search in large state spaces with high branching factors and an objective function to be maximized. Our method portfolio, which we refer to as heuristically guided swarm tree search, is randomized, as it consists of several Monte-Carlo runs, and guided, as it relies on fitness selection. We apply different search enhancement such as UCT, look-aheads, multiple runs, symmetry detection and parallel search to increase coverage and solution quality. Theoretically, we show that UCT, which trades exploration for exploitation, can be more successful on several runs than on only one. We look at two case studies. For the Same Game we devise efficient node evaluation functions and tabu color lists. For Morpion Solitaire the graph to be searched is reduced to a tree. We also adapt the search to the graphics processing unit.