Publications

Proposes an adaptive crossover selection mechanism that dynamically chooses between binomial and exponential crossover strategies based on historical performance, significantly improving DE's robustness across diverse problem classes on CEC2017 benchmarks.

Applies zero-shot learning with short-time Fourier transform features to classify radar and communication signals, demonstrating generalization to unseen signal types without retraining.

Introduces ModSHADE, a DE algorithm built on SHADE with triple-mutation ensemble, dual-bin crossover with adaptive rates, and probability-based parameter adaptation. Tested on CEC 2014 and 2017 benchmarks, it achieves improved convergence speed and robustness over standard SHADE.

A comparative study of state-of-the-art adaptive and hybrid DE variants, analyzing their performance on CEC 2017 benchmark functions. Highlights competitive performance in high-dimensional and complex scenarios, and outlines future directions for large-scale optimization and AI integration.

Introduces MGO, a fundamentally new metaheuristic paradigm where the population is a weighted, dynamically evolving graph rather than independent vectors. Five biologically grounded mechanisms—hyphal tip growth, Laplacian nutrient translocation, anastomosis crossover, cord formation, and spectral-entropy sporulation—replace standard operators, with provable spectral properties.

Introduces DECAS, a differential evolution algorithm for multi-cloud workflow scheduling with a cloud-aware repair operator, a cloud-bias mutation term, and a KNN-based surrogate that cuts simulation time by 50–80%. Across six benchmark workflows it delivers 28–35% cost savings over HEFT and 12–18% over IC-PSO, with Friedman-test significance at p < 0.001.

Introduces boundary decision consistency, a measure of how often a surrogate correctly orders solution pairs separating selected from rejected individuals, and argues this—not prediction accuracy—controls how well optimization proceeds. Proposes decision-adaptive surrogate training that directs learning toward pairs near the selection threshold. Across nine benchmarks and three practical applications, it beats both standard surrogate-assisted and ranking-based methods, with boundary consistency forecasting optimization quality far better than prediction error.

Establishes the first rigorous theoretical foundation for evolutionary optimization in cloud resource allocation—proving Multi-Objective Cloud Resource Allocation is strongly NP-hard with a 3/2 vector-packing inapproximability threshold, and that First-Fit Decreasing can be arbitrarily suboptimal. Introduces Ecra, a permutation-coded genetic algorithm with capacity repair, proven via multiplicative drift analysis to attain an O(log n)-approximation in expected polynomial time.

Proposes RADAR, a Risk-Aware Decision Architecture for autonomous serverless resource management that puts uncertainty quantification and risk budgeting at its center—letting operators specify their acceptable SLA violation probability via a single interpretable parameter. Combining conformal prediction, online distribution shift detection, and a risk-budgeted decision gate, RADAR cuts SLA violations by up to 72% during distribution shifts and achieves Pareto-dominant cost-safety tradeoffs over fixed-buffer baselines on Azure, Google, and Alibaba traces.

Designs and evaluates an IoT-based smart medication reminder system that improves patient adherence through automated alerts, real-time monitoring, and caregiver integration.