Hybridized self adaptive memetic optimization for green cloud scheduling using cross domain adaptation
The imperative for Green AI has intensified the need for intelligent cloud schedulers that can minimize carbon emissions without violating business-critical Service Level Agreements (SLAs). While native discrete metaheuristics have been proposed, they often fail to leverage the full power of state-of-the-art continuous optimizers. This paper introduces a dual contribution. First, we propose a new continuous optimizer, the Self-adaptive memetic optimizer (SA-MO), designed for high performance with architectural principles from elite competition-winning algorithms. It integrates a hybridized global search engine, combining operators inspired by Differential Evolution and Particle Swarm Optimization, a memetic local search for intense exploitation, and a self-adaptive mechanism for dynamic parameter control. The core SA-MO engine is first validated against elite continuous optimizers, including L-SHADE and CMA-ES, on ten standard benchmark functions, where it demonstrates superior convergence and precision. Second, we introduce a novel cross-domain adaptation framework to apply this powerful continuous optimizer to the discrete, combinatorial domain of cloud scheduling. The framework utilizes a priority-based encoding and decoding scheme that serves as a high-fidelity bridge. The generality of this framework is verified by successfully adapting standard continuous optimizers to the discrete domain. The core SA-MO engine is first validated against elite continuous optimizers, including L-SHADE and CMA-ES, on ten standard benchmark functions, where it demonstrates competitive convergence and precision within the evaluation budget used in this study. The resulting scheduler, d-SA-MO, is then benchmarked against industry heuristics and state-of-the-art native discrete metaheuristics across six challenging operational scenarios on a bi-objective formulation (carbon emissions and SLA violations), with energy consumption reported as a secondary monitoring metric. Quantitative results are presented in a condensed table, and rigorous non-parametric statistical analysis confirms that d-SA-MO achieves statistically significant superiority in overall performance (p < 0.0001). A complete ablation study and parameter sensitivity analysis scientifically validates the necessity of each component of the novel SA-MO architecture. This work presents a complete, end-to-end contribution: a new continuous algorithm and a novel framework to apply it, demonstrating a highly effective new pathway for solving complex Green AI challenges.