ICBS: Improved Conflict-Based Search Algorithm for Multi-Agent Pathfinding

Conflict-Based Search (CBS) and its enhancements, Meta-Agent CBS and bypassing conflicts are amongst the strongest newly introduced algorithms for Multi-Agent Path Finding. This paper introduces two new improvements to CBS and incorporates them into a coherent, improved version of CBS, namely ICBS. Experimental results show that each of these improvements further reduces the runtime over the existing CBS-based approaches. When all improvements are combined, an even larger improvement is achieved, producing state-of-the art results for a number of domains. Introduction and Overview A Multi-Agent Path Finding (MAPF) problem is defined by a graph, G = (V,E) and a set of k agents labeled a1 . . . ak, where each agent ai has a start position si ∈ V and a goal position gi ∈ V . At each time step an agent can either move to an adjacent location or wait in its current location. The task is to plan a sequence of move/wait actions for each agent ai, moving it from si to gi such that agents do not conflict, i.e., occupy the same location at the same time, while aiming to minimize a cumulative cost function. MAPF has practical applications in video games, traffic control, robotics etc. (see Sharon et al. 2013; 2015 for a survey). In this paper we focus on solving MAPF problems optimally, i.e., where the cost of the resulting plan must be minimized. There is a range of algorithms that optimally solve different variants of MAPF using various search techniques [Standley, 2010; Wagner and Choset, 2011; Sharon et al., 2013] or by compiling it into other known NP-hard problems [Surynek, 2012; Yu and LaValle, 2013; Erdem et al., 2013]. Each of these solvers has pros and cons, with no universal winner. Which algorithm performs best under what circumstances is an open research question. Conflict-Based Search (CBS) [Sharon et al., 2012a; 2015], is a very effective optimal MAPF solver. CBS has two-levels. The low-level finds optimal paths for the individual agents. If the paths include conflicts, the high level, via a split action (described below), imposes constraints on the conflicting agents to avoid these conflicts. Two improvements to CBS were suggested. First, Metaagent CBS (MA-CBS) [Sharon et al., 2012b; 2015] generalizes CBS by merging small groups of agents into meta-agents when beneficial. The main merge policy is to merge agents for which the number of conflicts seen so far exceeds a given threshold B; for good choices of B, MA-CBS reduces the runtime. Second, the bypass (BP) improvement to (MA)CBS (i.e., CBS with or without applying the optional merge action) was recently suggested [Boyarski et al., 2015]. (MA)CBS arbitrarily chooses paths in the low-level. When BP is added on top of (MA)CBS we first try to find an alternative path (bypass) for one of the conflicting agents thus avoiding the need to perform a split and to add new constraints. (MA)CBS+BP further reduces the running time. A variant of CBS that solves MAPF suboptimally also appeared [Barer et al., 2014]. In this paper we introduce Improved-CBS (ICBS) which further adds two new improvements to (MA)CBS+BP. Each of these improvements is strongly tied to one of the former improvements (MA and BP) as follows: • (1) Merge and restart. (MR) In MA-CBS, agents are merged locally at each node of the search tree. Instead, when a decision to merge is made, we suggest to restart the search from scratch, with the new merged meta-agent treated as a single agent for the entire search tree. • (2) Prioritizing conflicts. (PC) (MA)CBS (even with BP added) arbitrarily chooses which conflict to split. Poor choices may increase the size of the high-level search tree. To remedy this, we prioritize the conflicts according to three types: cardinal, semi-cardinal and non-cardinal. Cardinal conflicts always cause an increase in the solution cost, so ICBS chooses to split cardinal conflicts first. Additionally, bypasses to cardinal conflict cannot exist. Therefore, the optional BP improvement [Boyarski et al., 2015] should only be applied for semior non-cardinal conflicts. All four enhancements to CBS (i.e., MA-CBS, BP, PC and MR) are optional and can be added separately or in conjunction with the others, except for MR which is only relevant to MA-CBS. In this paper we show how to combine them all into a coherent improved version of CBS, namely ICBS. Experimental results show that MR and PC each provide a significant speedup over the previous best CBS variant, namely (MA)CBS+BP [Boyarski et al., 2015]. Even further speedup is achieved when both improvements are combined and ICBS outperforms other related algorithms in many cases where previous variants of CBS performed poorly. Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence (IJCAI 2015)