In a development that could reshape how modern factories juggle speed and sustainability, a team of researchers in China has unveiled a bio-inspired artificial intelligence system that teaches itself to orchestrate complex manufacturing schedules. The new approach, published in the journal Applied Intelligence, tackles one of the most notoriously difficult problems in industrial engineering: deciding, second by second, which machines in a network of distributed factories should process which jobs, in what order, and with how much idle energy expenditure. The work, led by Mengxin Tao and colleagues at the Qingdao University of Science and Technology, combines the foraging behavior of slime mold with a reinforcement learning technique borrowed from game-playing AI to simultaneously minimize production time and energy consumption.
The problem the researchers set out to solve is formally known as the Distributed Assembly Hybrid Flow Shop Scheduling Problem with Sequence-Dependent Setup Times, abbreviated DAHFSP-SDST. Its structure reflects the reality of contemporary manufacturing, where products are rarely built in a single facility. Instead, component parts are fabricated across several geographically distributed factories, each of which contains multiple stages of production, and each stage may operate several parallel machines. Once the components are complete, they are shipped to assembly facilities where they are combined into finished products. Complicating matters further, when a machine switches from producing one type of component to another, it often requires a setup time that depends on the sequence of the tasks — switching between similar products may be quick, while a changeover between radically different ones may be slow and energy-intensive. This sequence-dependent setup time introduces a combinatorial explosion: the cost of scheduling one job depends on which job came before it, meaning that even modestly sized problems can contain more possible schedules than atoms in the observable universe.
Scheduling problems of this class are classified as NP-hard in the computational complexity sense, which in practical terms means that no known algorithm can guarantee finding the perfect answer in reasonable time as the problem grows. Factories therefore rely on heuristics and metaheuristics — clever search strategies that sacrifice the guarantee of perfection in exchange for very good answers found quickly. The Qingdao team’s contribution, called the Improved Multi-Objective Slime Mold Algorithm, or IMOSMA, belongs to this family but introduces several innovations that allow it to navigate the vast search space with unusual efficiency while balancing two competing objectives at once: makespan, the total time needed to complete all products, and total energy consumption.
The foundation of the algorithm is the slime mold algorithm, a stochastic optimization method introduced in 2020 that draws inspiration from the behavior of Physarum polycephalum, a single-celled organism famous for its ability to find efficient paths through mazes and build near-optimal transport networks without any central brain. In the algorithmic analog, candidate solutions behave like colonies of slime mold, spreading through the space of possible schedules, attracted toward promising regions and thinning out in barren ones. While the slime mold algorithm has proven effective on continuous optimization problems, adapting it to discrete scheduling — where solutions are permutations and assignments rather than smooth numerical vectors — is a significant challenge, and this is where the researchers’ improvements come into play.
The first innovation concerns how the algorithm starts its search. A population-based optimizer’s fate is largely sealed by its initial solutions: if the starting population is poor or too uniform, the search converges prematurely on mediocre schedules. The team designed a hybrid initialization strategy that blends three ingredients. Chaotic mapping replaces the standard pseudo-random numbers used to seed the population with deterministic sequences generated by chaotic dynamical systems, which wander through their range in a way that is more thoroughly distributed than pure randomness. Gaussian perturbation adds small, statistically distributed tweaks to the initial solutions, diversifying them further. Finally, the Longest Processing Time rule — a classic scheduling heuristic that assigns the longest jobs first, mimicking the strategy of loading the heaviest suitcases into a trunk before filling gaps with small ones — injects a handful of high-quality, experience-informed seeds into the population. The combination gives the algorithm a population that is simultaneously diverse and well-informed, setting the stage for rapid early progress.
Once the search begins, IMOSMA employs what the authors call a dual-population co-evolution mechanism. Rather than evolving one homogeneous swarm of candidate schedules, the algorithm splits its population into two cooperating groups. An elite population, comprising the best solutions found so far, undergoes position-based crossover, an operator adapted from evolutionary computation that exchanges structural fragments between high-performing schedules to produce offspring that inherit the strengths of both parents. The normal population, meanwhile, learns directly from these elite solutions, updating its own members by moving toward the elite positions according to the slime mold’s characteristic oscillatory dynamics. This division of labor accelerates convergence: the elite group refines the best-known regions of the search space while the general population continually harvests that progress, a feedback loop that prevents either group from stagnating.
Perhaps the most striking element of the system, however, is its use of reinforcement learning to steer the local search. Local search — repeatedly making small changes to a schedule to see if it improves — is essential in modern metaheuristics, but the effectiveness of any particular change operator varies enormously depending on the current state of the solution. A move that rescues a bottlenecked factory floor might be useless elsewhere. The researchers equipped IMOSMA with a Double Deep Q-Network, or DDQN, a deep reinforcement learning architecture that originated in the Atari-playing agents of DeepMind. The DDQN module has eight neighborhood operators at its disposal, each representing a different way of perturbing a schedule. At each step of the search, the network observes features of the current state — including the characteristics of production bottlenecks and the balance between the time and energy objectives — and selects the operator it has learned to be most promising. Because it is a “double” deep Q-network, the architecture uses two separate neural networks to decouple action selection from action evaluation, a well-known trick that mitigates the overestimation bias that can destabilize standard Q-learning. Over the course of many scheduling episodes, the network learns a dynamic policy: which kind of local move works best in which kind of search situation.
The overall effect is an algorithm that behaves less like a fixed search procedure and more like an adaptive artisan — probing the schedule with different moves, recognizing which tools suit the moment, and allocating its search effort where it matters most for both objectives. This matters because makespan and energy consumption frequently pull in opposite directions. Compressing a schedule may force machines to run at full power with frequent changeovers, inflating energy use, while an energy-lean schedule may let machines idle or power down, stretching the timeline. Multi-objective optimization handles this tension by seeking a Pareto front: a set of schedules in which no solution can be improved on one objective without worsening the other. The goal is not a single perfect schedule but a rich menu of trade-off options from which a plant manager can choose depending on whether delivery deadlines or electricity costs dominate on a given day.
To test their method, the researchers generated a battery of benchmark instances following established procedures in the scheduling literature, spanning a range of problem sizes and configurations. They compared IMOSMA against a lineup of competitive algorithms from the published state of the art. The results were consistent: IMOSMA outperformed the comparative algorithms on the majority of test instances, delivering higher-quality Pareto fronts with more stable performance across runs. The authors attribute this robustness to the synergy of their three mechanisms — the diverse initialization gives the search a strong and varied start, the dual-population co-evolution sustains momentum through the middle of the search, and the DDQN-guided operator selection prevents the local search from wasting effort on unproductive moves, particularly in the delicate final phase where both time and energy objectives must be fine-tuned simultaneously.
The implications extend beyond academic benchmarks. Distributed manufacturing — in which production is spread across multiple plants, often to be close to suppliers, customers, or specialized expertise — is a cornerstone of modern supply chains in industries ranging from automotive to electronics. Energy costs and carbon accounting have moved to the center of industrial strategy, and regulatory frameworks increasingly demand that efficiency gains come not just from hardware upgrades but from smarter operation of existing equipment. An algorithm that can shave hours off a production calendar while simultaneously trimming kilowatt-hours addresses both concerns with a single software intervention. Because the approach learns its operator selection policy from experience, it also holds promise for transfer across different factory configurations, although the authors note that further validation on industrial case data would be a natural next step.
The research also exemplifies a broader trend in artificial intelligence: the fusion of nature-inspired optimization with machine learning, sometimes called learning-assisted metaheuristics. Rather than choosing between evolutionary computation and reinforcement learning, the field increasingly blends them, letting swarms explore broadly while learned policies make the fine-grained decisions. The Qingdao team’s slime mold-plus-DDQN hybrid is a vivid illustration of this convergence, with an organism that has no brain at all providing the architectural metaphor for a system smart enough to optimize the nervous system of a factory. The work was supported by the National Natural Science Foundation of China under Grant No. 62201314, and the authors have made their benchmark instance generation procedures and code available to other researchers on reasonable request, paving the way for the wider community to build on their results.
Subject of Research: Multi-objective optimization of the distributed assembly hybrid flow shop scheduling problem with sequence-dependent setup times, using an improved slime mold algorithm enhanced with DDQN-based reinforcement learning to minimize makespan and energy consumption.
Subject of Research: Technology and Engineering
Article Title: Distributed assembly hybrid flow shop scheduling optimization based on the slime mold algorithm and reinforcement learning
Article References: Tao, M., Wei, X., Li, X., & Zhou, Y. (2026). Distributed assembly hybrid flow shop scheduling optimization based on the slime mold algorithm and reinforcement learning. Applied Intelligence, 56(13), Article 389. https://doi.org/10.1007/s10489-026-07419-3
Image Credits: AI Generated
DOI: 10.1007/s10489-026-07419-3
Keywords: Distributed assembly hybrid flow shop scheduling, multi-objective optimization, slime mold algorithm, DDQN, reinforcement learning, sequence-dependent setup times, makespan, energy consumption, metaheuristics, intelligent manufacturing
Cite Scienmag News
APA
MLA
Chicago
Denise Maddox. (September 6, 2026). Slime mold algorithm meets reinforcement learning to optimize distributed assembly scheduling. Scienmag. https://scienmag.com/slime-mold-algorithm-meets-reinforcement-learning-to-optimize-distributed-assembly-scheduling/
Denise Maddox. “Slime mold algorithm meets reinforcement learning to optimize distributed assembly scheduling.” Scienmag, 6 September 2026, https://scienmag.com/slime-mold-algorithm-meets-reinforcement-learning-to-optimize-distributed-assembly-scheduling/. Accessed 6 September 2026.
Denise Maddox. “Slime mold algorithm meets reinforcement learning to optimize distributed assembly scheduling.” Scienmag. September 6, 2026. https://scienmag.com/slime-mold-algorithm-meets-reinforcement-learning-to-optimize-distributed-assembly-scheduling/
Copy citation
Download RIS
Tags: adaptive reinforcement learning in industrial engineeringAI-driven factory orchestrationAI-driven industrial schedulingbio-inspired AI for factory optimizationbio-inspired AI for industrial optimizationcomplex industrial schedulingcomplex manufacturing process optimizationdistributed assembly schedulingdistributed manufacturing systemsenergy-efficient manufacturinghybrid flow shop schedulingmulti-factory production managementmulti-factory production planningoptimization of production time and energyreinforcement learning in manufacturingsequence-dependent setup timessequence-dependent setup times in productionSlime mold algorithmsustainable factory operations
