On Improving the Quality of Solutions in Large-Scale Cooperative Multi-Agent Pathfinding

Scaling up the number of simultaneously moving units in pathfinding problems to hundreds, or even thousands, is well beyond the capability of theoretically optimal algorithms in practice, which is consistent with existing intractability results (Surynek 2010). However, significant scalability can be achieved by trading off solution optimality. This motivates evaluating the quality of suboptimal solutions, especially in instances much larger than can be handled by optimal algorithms. In this work, we consider pathfinding in uniform cost grid maps. We study the solution quality using the three most common quality criteria. The total travel distance is the sum of distances covered by all units, measuring the total cost. The sum of actions includes moves and also wait actions that appear in a plan, providing an indication of the aesthetic quality of a plan, as long or frequent wait intervals are undesirable. The makespan measures the total number of time steps in a global plan when steps can be run in parallel. We focus on MAPP (Wang and Botea 2009; 2010), which has been shown as state-of-the-art in terms of scalability and success ratio (i.e., percentage of solved units) on realistic game grid maps. Until now, the quality of MAPP’s solutions had not been as extensively analyzed. Our analysis indicates that solutions computed with existing versions of MAPP can often suffer in terms of quality. We introduce enhancements that significantly improve MAPP’s solution quality. For example, the sum of actions is cut to half on average. MAPP becomes competitive in terms of solution quality with FAR (Wang and Botea 2008) and WHCA* (Silver 2005), two successful algorithms from the literature. MAPP provides a formal characterization of problems it can solve, and low-polynomial upper bounds on the resources required, which are not reported for FAR and WHCA*. Our extensions bring MAPP’s solution quality to a state-of-theart level, while maintaining its advantages over FAR and WHCA* on the performance criteria of scalability, success ratio, and ability to tell a priori if it will succeed in the instance at hand. One of our extensions, which spreads out the