population management

This work proposes a population-based neural combinatorial optimization framework that addresses the limited exploratory capacity and robustness of traditional neural approaches, which typically operate on a single solution. By leveraging neural networks to jointly represent a set of candidate solutions, the framework incorporates a population-aware hierarchical classification mechanism to explicitly model inter-solution information sharing and diversity control. This design simultaneously reinforces high-quality solutions and preserves population diversity, effectively bridging the gap between neural optimization and classical population-based metaheuristics. Experimental results on the Max-Cut and Maximum Independent Set problems demonstrate that the proposed framework substantially improves both solution quality and algorithmic robustness.

In service-oriented organizations, personnel scheduling, team management, and position deployment constitute three tightly coupled human-resource optimization problems; existing approaches typically model them in isolation, failing to handle dynamic, non-stationary environments. Method: We propose the first unified multi-scale labor optimization simulation framework—featuring a modular, configurable simulation environment; the first integration of long-horizon performance objectives into a multi-agent deep reinforcement learning (MARL) framework for cross-temporal-scale joint optimization; and the joint incorporation of dynamic stochasticity modeling and heuristic baseline integration. Contribution/Results: Our method achieves significant improvements on standardized benchmarks: +18.3% in long-term labor efficiency, −22.7% in response latency, and +15.1% in resource utilization. It supports ablation studies and policy generalization evaluation, thereby bridging critical gaps in unified modeling of interdependent decisions and MARL-driven long-term optimization.

Existing population-based training methods for zero-shot coordination (ZSC) suffer from prohibitive computational costs and poor scalability with population size. To address this, we propose ScaPT, a scalable population training framework that models large cooperative populations via parameter-sharing meta-agents and incorporates mutual information regularization to preserve behavioral diversity while drastically reducing computational overhead. Experiments on the Hanabi benchmark demonstrate that ScaPT significantly outperforms prior approaches in zero-shot coordination performance. Notably, it provides the first empirical validation that increasing population size yields substantial gains in ZSC generalization capability. Moreover, ScaPT exhibits strong scalability—maintaining efficiency even as population size grows. This work establishes a novel paradigm for efficient and scalable multi-agent cooperative learning, advancing the state of the art in population-based ZSC.

In zero-shot coordination (ZSC), two core bottlenecks impede progress: insufficient partner agent diversity and inefficient cross-game policy minimization (XPM) training—XPM relies on costly multi-trajectory environment sampling and requires independent, from-scratch training for each partner. To address these, we propose XPM-WM, the first framework to integrate a world model (comprising a VAE and RSSM) into XPM. It replaces expensive real-environment trajectory sampling with model-generated synthetic trajectories, eliminating the need for multi-trajectory collection. Moreover, a shared, reusable dynamics model drives the evolution of diverse partner policies, obviating redundant per-partner training. Evaluated on the SP benchmark, XPM-WM matches state-of-the-art performance in ZSC success rate and population training reward while improving sample efficiency by 3.2× and enabling efficient generation of partner agents at scale (up to hundreds).

This work addresses the challenge of response mapping failure in large-scale multi-agent systems caused by dynamic population changes, which traditional planners struggle to handle. The authors propose a population-aware coordination interface grounded in a Lagrangian relaxation framework, employing conditional neural networks to learn primal and dual mappings from compact population summaries. This enables stable cross-episode predictions without retraining, effectively adapting to evolving populations. The approach efficiently coordinates massive agent populations (e.g., 500,000 agents) using only small subsamples (e.g., 20,000 agents) and reformulates Sim2Real transfer into a backtestable pipeline. Evaluated on a supply chain capacity control task, the method reduces prediction error by 16–19% and capacity violations by 20–51% compared to baselines. When applied to real-world data, the simulation-trained model achieves a MAPE of 11.1%, substantially outperforming baseline methods reporting 13–24%.

Traditional evolutionary game analysis relies on closed-form payoff expressions derived from queueing systems, which are infeasible in complex scenarios. This work proposes a Discrete-Event Population Update (DEPU) framework that directly embeds a single discrete-event simulation into the evolutionary dynamics, enabling efficient analysis of strategy evolution in systems lacking closed-form solutions without resorting to nested simulations. The framework encompasses two implementation mechanisms: Discrete-Event Replicator Dynamics (DERD) and Discrete-Event Moran Replacement (DEMR). Evaluated on a multi-server jockeying queue model, DEPU achieves computational speedups of an order of magnitude over conventional methods while preserving comparable accuracy, substantially enhancing the feasibility and efficiency of large-scale parameter sweeps.

This study investigates whether the resource cycling mechanism in Metabolic Multi-Agent Optimizers (MMAO) possesses framework-level explanatory power beyond metaphorical interpretation. By constructing an abstract state model that strips away domain-specific details while preserving the core resource accounting mechanism, the work leverages abstract modeling, dynamical analysis, and stability theory under mild bounded gain-and-expenditure assumptions to reveal, for the first time, the endogenous regulatory dynamics of MMAO at the architectural level. The analysis establishes the non-negativity and boundedness of key resource variables—such as private energy and public budget—and identifies three universal behavioral modes: contraction, reinvestment, and search reallocation. The generality of these modes is validated through both continuous and discrete instantiations, clearly distinguishing behaviors inherent to the metabolic feedback loop from those contingent on implementation specifics.

This work proposes a novel framework for jointly optimizing the skills of large language model (LLM) agents, where each skill comprises instructions, tools, and supporting resources whose structure and content are tightly coupled, resulting in a complex and interdependent optimization landscape. For the first time, this joint optimization is formalized as a bilevel problem: the outer loop employs Monte Carlo Tree Search to explore skill structures, while the inner loop optimizes the content of components given a fixed structure, with both loops leveraging LLMs to guide decision-making. By integrating structured skill representations with an efficient search strategy, the approach significantly enhances agent performance on an open-source operations research question-answering benchmark, demonstrating both the effectiveness and novelty of the proposed framework.

This work addresses the lack of a unified convergence analysis framework for population-based optimization algorithms, which hinders systematic comparison and generalization. The authors propose an operator calculus framework that models diverse algorithms as compositions of three fundamental operators—mutation, selection, and recombination—acting on probability measures. By leveraging mean-field limits, they derive a continuous-time transport-reaction-jump partial differential equation governing the algorithmic dynamics. Building upon operator semigroup theory and functional analysis on spaces of probability measures, they develop a modular Lyapunov method that enables dissipativity verification operator by operator. Under explicit stability and regularity conditions, they establish exponential decay of both a state-space Lyapunov functional and the search error, thereby providing a unified guarantee of exponential convergence for a broad class of distributed optimization algorithms.