New Self-Adaptive Optimizer Bridges the Continuous-Discrete Divide to Green Cloud Scheduling

The digital world runs on data centers, and data centers run on electricity. As artificial intelligence workloads surge into the exascale era, the Information and Communication Technology sector has accumulated a carbon footprint so large and fast-growing that it now threatens global sustainability targets. A new open-access study published in Discover Informatics proposes a strikingly original answer to this problem: a scheduling algorithm that smuggles the raw power of elite continuous optimization techniques into the stubbornly discrete world of cloud task assignment, cutting carbon emissions while still honoring the service-level agreements that businesses cannot afford to breach. The result, its authors argue, is a complete end-to-end pathway for solving one of Green AI’s hardest problems.

The work, led by Raza Hasan of Southampton Solent University together with Salman Mahmood, Sellappan Palaniappan, and Deborah Adedigba, makes a dual contribution. The first half is a brand-new continuous optimizer called the Self-adaptive Memetic Optimizer, or SA-MO. The second half is a cross-domain adaptation framework that lets this real-valued algorithm tackle the combinatorial puzzle of assigning tasks to servers, producing a scheduler the team calls d-SA-MO. The core insight is simple but provocative: decades of progress in continuous numerical optimization, exemplified by competition-winning algorithms from the IEEE Congress on Evolutionary Computation, has largely gone untapped by scheduling researchers because the two domains speak different mathematical languages.

SA-MO itself is a synthesis of three architectural principles drawn from the state of the art in evolutionary computation. Its global search engine hybridizes Differential Evolution with Particle Swarm Optimization, dynamically choosing between a diversity-preserving DE/rand/1 exploratory operator, which builds new candidate solutions from the vector differences of randomly selected population members, and an exploitative PSO-style operator that pulls individuals toward personal-best and global-best positions. A linearly decaying switch probability, starting at 0.9 and falling to 0.1, ensures the search begins with broad exploration of the fitness landscape and gradually intensifies toward refinement. Layered on top is a memetic local search routine based on Powell’s Conjugate Direction method, applied stochastically to the top performers in the trial population. Because the decoded scheduling objective behaves as a black-box function with no analytically computable gradient, Powell’s method was chosen precisely because it does not require derivatives. Finally, borrowing from Evolution Strategies, each individual carries an array of strategy parameters, or mutation step sizes, that are themselves mutated and evolved, allowing the algorithm to learn the most effective search strengths for whatever landscape it faces.

Before touching a single cloud workload, the team validated the SA-MO engine on ten standard continuous benchmark functions against two elite competitors: L-SHADE and CMA-ES. Under the evaluation budget matched to the scheduling task, SA-MO achieved the lowest mean error on several functions, including the Sphere, Schwefel, and Dixon-Price problems, and the authors characterize the engine as competitive with, rather than definitively superior to, the CEC champions. That honesty matters; the point of the exercise was not to break convergence records but to demonstrate that the new engine belongs in the same tier as the field’s best before being entrusted with an NP-hard scheduling problem.

The bridge from continuous to discrete is where the paper’s novelty concentrates. Continuous optimizers manipulate real-valued vectors in n-dimensional space, while cloud scheduling is a binary assignment problem: each of n tasks must go to exactly one of k heterogeneous, geographically distributed servers. Earlier attempts to cross this divide have well-documented flaws. Naive rounding destroys the gradient information continuous optimizers depend on and often produces infeasible schedules requiring repair. Random-key encodings have mostly been confined to simplified single-objective problems like traveling-salesman variants. Purpose-built discrete metaheuristics operate natively on schedules but cannot inherit two decades of continuous-optimization innovation. The new framework sidesteps all of this with a priority-based encoding and decoding scheme. Each solution inside the optimizer is a continuous vector of task priorities; a deterministic decoder ranks the tasks and greedily assigns them to servers, respecting capacity constraints and producing feasible schedules every time. The optimizer thus treats the discrete problem as a black-box continuous function, its memetic and self-adaptive machinery untouched.

The scheduling problem itself is formalized as a bi-objective optimization. Because energy consumption and carbon emissions are related by a simple multiplicative factor, the team optimizes carbon directly, weighting each server’s energy use by the time-variable carbon intensity of its local electricity grid, modeled from real-world 24-hour traces provided by the ElectricityMaps API. Minimizing this objective rewards routing workloads toward regions and time windows where renewables dominate the energy mix. The second optimized objective is SLA satisfaction, with violations triggered when total latency, processing time plus network delay, exceeds each task’s contractual threshold. Energy consumption is reported as a secondary monitoring metric. The two competing objectives are fused into a single Composite Score, normalized by a scaling constant set at one million, an order-of-magnitude calibration confirmed by sensitivity analysis.

The empirical campaign was extensive: seven algorithm variants, six operational scenarios ranging from a low-carbon Green Grid to high-contention environments, and 30 independent runs each, totaling 1,260 optimization runs on problem instances of 500 tasks and 20 heterogeneous servers. The headline result is unambiguous. d-SA-MO achieved the best mean Composite Score, the lowest carbon emissions, and the fewest SLA violations in every single scenario, with the tightest variance. A Friedman omnibus test rejected equality of algorithms with a p-value of roughly one times ten to the minus 23, and pairwise Wilcoxon signed-rank tests with Holm-Bonferroni correction confirmed the advantage over the strongest native discrete baseline, a state-of-the-art memetic algorithm called MAJO, at p below 1.12 times ten to the minus 8. Effect sizes were equally emphatic: in the High Contention scenario, Cliff’s delta reached 0.94 against MAJO, meaning d-SA-MO won 94 percent of paired comparisons, and the advantage stayed in the very large effect band across all six environments.

An ablation study dissects why the architecture works. Removing the memetic local search dropped the mean score from 65.35 to 62.76 in the High Redundancy scenario, a statistically significant loss proving that global operators alone lack the precision of hybridized exploitation. Disabling self-adaptation was even more revealing: performance barely changed in the benign Green Grid setting but collapsed by more than five points in the rugged Dirty Grid landscape, demonstrating that adaptive parameter control is the algorithm’s insurance policy against harsh environments. Replacing the hybrid engine with a standard genetic search consistently underperformed the full model. The paper also established that the framework itself, not just the engine, is general-purpose: when the same priority-based bridge was fitted to standard Differential Evolution and Particle Swarm Optimization, the resulting schedulers d-DE and d-PSO both surpassed the native discrete MAJO, with median Composite Scores near 78 and 76 respectively against roughly 72.

Practicality was tested through scalability experiments scaling from 100 tasks and 10 servers to 2,000 tasks and 50 servers. The theoretical complexity of O(Gmax times Npop times (n log n + nk)) translated into polynomial runtime growth in practice, and remarkably, the adapted continuous algorithm actually outpaced its discrete rival: at the largest scale, d-SA-MO finished in 99.05 plus or minus 10.07 seconds versus 165.58 plus or minus 18.63 for MAJO, a difference the authors report as statistically significant. Solution quality degraded only slightly with scale, with d-SA-MO maintaining a persistent lead of about 3.3 to 3.5 points over MAJO at every instance size. Sensitivity analyses rounded out the validation, identifying p equal to 0.2 as the sweet spot for the local-search trigger probability and confirming, through a cross-evaluation procedure, that solutions trained under an overly large scaling constant mask their true carbon cost when judged at the canonical scale.

The implications reach beyond one scheduler. By proving that the supposed domain mismatch between continuous and discrete optimization is structural rather than inherent, the study opens a general route for deploying elite optimizers on combinatorial Green AI challenges, a result with direct relevance to United Nations Sustainable Development Goals on affordable and clean energy, industry and innovation, responsible consumption, and climate action. The authors note their simulator remains an abstraction, and they point to concrete next steps: extending SA-MO to true Pareto-based multi-objective frameworks such as NSGA-II and MOEA/D, validating the scheduler on a physical Kubernetes testbed, and exploring learned decoders that could replace the handcrafted priority bridge. For now, the message to cloud operators is clear: the most powerful mathematical engines ever built for optimization no longer need to stay in continuous space, and the climate may be the beneficiary.

Subject of Research: Green cloud scheduling using a hybridized self-adaptive memetic optimizer adapted from continuous optimization via a cross-domain encoding framework

Article Title: Hybridized self adaptive memetic optimization for green cloud scheduling using cross domain adaptation

Article References: Hybridized self adaptive memetic optimization for green cloud scheduling using cross domain adaptation. (n.d.). https://doi.org/10.1007/s44564-026-00005-2

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00005-2

Keywords: green cloud computing, carbon-aware scheduling, memetic algorithm, differential evolution, particle swarm optimization, self-adaptation, cross-domain adaptation, service level agreements, metaheuristics, multi-objective optimization, Green AI, data center emissions