Final-year BS Computer Science at NUST SEECS, Islamabad.

I work on how optimization algorithms are evaluated. Each project below takes an experimental design that is standard in the metaheuristics literature and adds the one control it usually leaves out: a random baseline over the same operators, the exact optimum, a held-out test fold, a full budget profile. In all four, adding the control changes the conclusion.

None of these effects is new on its own, and each repository cites the work it reproduces. What the projects add is a measurement of how large each effect is, under one statistical protocol.

Shared across all four: evaluation budgets enforced in code, seeds recorded per run, Friedman tests with Holm-corrected Wilcoxon post-hoc comparisons and Cliff's δ effect sizes, and a stated Limitations section. Every table and figure regenerates from the raw results with one make target. MIT licensed.

Five optimizers (DE, PSO, GA, IPOP-CMA-ES, random search) on the 24 BBOB functions at D=10 and 20,000 evaluations, plus a structural-bias test on an objective that carries no information.

• On a pure-noise objective, DE/rand/1/bin finishes biased toward the domain boundary in 10 of 10 dimensions after Holm correction, 15% farther from the centre than uniform sampling. Random search, the control, is not flagged. This reproduces the structural-bias effect described by Kononova et al. (2015).
• PSO solves 1.88× as many targets as DE at about 600 evaluations. DE overtakes it at about 5,200 and solves 1.76× as many by 20,000. A single end-of-run ranking hides the reversal.

Limits: default parameters, one dimension. Bias is measured for each algorithm together with its bound-handling rule, and this design cannot separate the two.

Five bandit policies choosing among five DE mutation operators, compared with uniform random selection over the same pool and with JADE. 24 BBOB functions, D=5 and D=20, 14,400 runs.

• At D=5 all five policies beat uniform random selection, four of them significantly. At D=20 none do, and two are significantly worse.
• Every policy converges on DE/best/1, the operator that improves the parent most often and the worst of the five fixed operators by final error. The harder a policy commits to its top operator, the worse it ranks (Spearman ρ = 0.90).
• Against a fixed best/1 baseline, the same D=20 runs show all five policies significantly better. The baseline decides the headline. Karafotias et al. (2013) made this argument for parameter control, and Hanna et al. (2025) found the same result in program repair.

Limits: restricted budget (500 × D), one BBOB instance, F and CR held fixed.

Multilevel Otsu and Kapur thresholding solved exactly by an O(C·L²) dynamic program, with PSO, DE, GA and random search measured against the proven optimum. 28,000 runs on 14 images, 96 tests.

• The dynamic program returns the proven optimum in every case, in 3.4–7.8 ms. At six classes the metaheuristics reach it in 81% (PSO), 64% (DE) and 48% (GA) of runs. Exact solutions have been in print since Liao et al. (2001); what is usually missing is the comparison against them.
• It is slower than scikit-image's threshold_multiotsu at two and three classes, and 40,968× faster at six.
• PSNR, SSIM and FSIM disagree with Dice about which criterion is better only where the two segmentations differ by less than 0.005 Dice, which is the range in which papers use them to separate methods.

Limits: separable criteria and 1-D histograms only. The metric comparison uses synthetic ground truth.

NSGA-II and NSGA-III feature selection on a synthetic dataset and the colon microarray dataset (n = 62), against matched-budget random subsets, t-test with Benjamini–Hochberg, mRMR, LASSO, Model-X knockoffs and an oracle, under nested cross-validation.

• On the synthetic data, NSGA-II beats random subsets decisively on the inner-CV objective (hypervolume δ = +0.999) and not on held-out error (δ = +0.065, Holm p = 1.00). Reporting inner-CV error would have shown a large win that is not there, the selection-bias problem described by Ambroise & McLachlan (2002).
• The t-test filter used 8 evaluations to NSGA-II's 888 and ranked better on held-out error. Of ten planted features it recovered seven; NSGA-II recovered none.

Limits: two datasets, one budget (900 evaluations), one inner classifier.

Four Omitted Controls in Metaheuristic Optimization Experiments. A preprint that puts the four studies under one protocol and one argument. Link to follow.

Bone mineral density from plain radiographs. Final-year project in the BIOMISA group at NUST, supervised by Prof. Dr. Muhammad Usman Akram. A self-supervised Vision Transformer estimating bone mineral density from standard spinal X-rays, as a low-cost alternative to DEXA scanning, trained on paired radiograph, DEXA and blood-count data from two Islamabad hospitals. In progress.

I also work as a full-stack and applied LLM engineer, currently at Swix AI: agent platforms, retrieval pipelines and real-time multimodal systems. That code is private client work and is not here.

Tools: Python, NumPy, SciPy, pandas, scikit-learn, pymoo, moocore, IOHexperimenter, scikit-posthocs, scikit-image, PyTorch, MATLAB, LaTeX.

Contact: hammadarif564@gmail.com