SubMAPL beats policy-gradient baselines in dynamic multi-agent task allocation

A research team led by Jing Liu and five collaborators at academic institutions has introduced SubMAPL, a new algorithm for distributed task allocation in open multi-agent systems (MAS) — environments where agents can dynamically join or leave. The work, posted on arXiv (2608.14390), tackles the challenge of coordinating a changing set of agents whose collective utility is submodular (i.e., exhibits diminishing returns). Prior continuous relaxations of submodular set functions relied on independent Bernoulli sampling, which creates a mismatch with categorical action distributions used by agents. To fix this, the authors propose the partition multilinear extension (PME), a policy-based relaxation whose continuous support matches feasible actions under categorical distributions. They prove that marginal gains of the stage utility provide an unbiased estimator of the PME gradient, and that maximizing the PME is equivalent to maximizing the true stage utilities — a critical step for principled policy learning.

Building on PME, SubMAPL is a centralized-training, decentralized-execution method that employs KL-mirror updates to preserve categorical feasibility without Euclidean projections. During training, local marginal gains serve as stochastic PME gradients. To handle agent arrivals and departures, the authors introduce open policy migration and an open-system KL tracking variant for tabular-softmax policies. The theoretical analysis establishes a dynamic regret lower bound that accounts for environment openness and the gap between stage-wise and global optimal utilities. Simulations on a multi-agent coverage task demonstrate that SubMAPL consistently outperforms policy-gradient and online-learning baselines, suggesting a robust path for real-world deployments in swarm robotics, warehouse logistics, and autonomous surveillance where team membership is fluid.