DyMAB Improves Multi‑Agent Pathfinding via Adaptive Bandits

Multi‑agent path finding (MAPF) remains one of the most challenging combinatorial optimization problems in robotics and logistics. A popular approach to tackle large instances is adaptive large neighbourhood search (ALNS), which iteratively applies a set of neighbourhood operators—heuristics—to explore new solutions. In practice, ALNS relies on simple selector mechanisms such as roulette wheel or classic multi‑armed bandit (MAB) policies to decide which operator to use at each iteration. These selectors assume that an operator’s reward distribution is stationary; however, in real MAPF runs the best operator can shift as the search progresses, leading to suboptimal exploration and wasted computational effort. ¹

To address this limitation, Amath Sow, Daniel de Leng, Mariusz Wzorek, and Fredrik Heintz propose DyMAB—a dynamic multi‑armed bandit framework for neighbourhood selection. The core idea is to embed a non‑stationary bandit model directly into the ALNS loop. DyMAB maintains a windowed queue that records recent rewards from every heuristic, allowing it to detect changes in performance trends. When a previously underperforming operator starts yielding higher returns, DyMAB quickly shifts its focus toward it, ensuring that the solver always prioritizes the most effective neighbourhood. ¹

The authors provide a theoretical analysis showing that DyMAB achieves sub‑linear regret under standard assumptions about non‑stationarity. Over time, DyMAB’s cumulative loss compared with an oracle that knows the best operator remains bounded by a function growing slower than linearly, giving a formal guarantee of long‑term efficiency. ¹

Expert commentary highlights that DyMAB’s use of non‑stationary multi‑armed bandits directly addresses limitations inherent in traditional ALNS approaches used for MAPF. The sub‑linear regret bound suggests that DyMAB will outperform standard MAB policies over time—an advantage given its anytime nature—and aligns with recent work from researchers like Klaus Ritzkow demonstrating significant gains using similar adaptive strategies in other combinatorial optimisation problems.

In conclusion, DyMAB represents a concrete advance for AI planning systems that must adaptively choose among many possible moves while dealing with shifting reward landscapes. Its sub‑linear regret guarantee positions it as a promising tool for practitioners seeking robust, self‑adjusting heuristics without manual parameter sweeps. The publication by Sow et al adds an important piece to the ongoing effort to make large‑scale pathfinding both efficient and reliable.