Home

Welcome to the cnn-optimization-benchmark wiki!

🔬 Welcome to the CNN Optimization Benchmark Wiki

The CNN Optimization Benchmark Platform is a research-grade scientific software suite designed for empirically benchmarking and comparing metaheuristic optimization algorithms applied to Deep Convolutional Neural Network (CNN) compression.

---

🎯 Primary Research Objectives

• Standardized Fair Comparison: Evaluate 10 state-of-the-art metaheuristics on identical CNN architectures, datasets, and hardware targets.
• Multi-Objective Trade-offs: Measure the multi-dimensional frontier across Accuracy (%), Latency (ms), Model Size (MB), and Energy Consumption (Joules).
• Stochastic Rigor: Aggregate multi-run repetitions with Mean, Median, Variance, and 95% Confidence Intervals.
• Interactive Workbench: Dynamic real-time objective re-weighting with live ranking recalculation.

---

🔬 Supported Metaheuristic Optimizers

1. GWO — Grey Wolf Optimizer (Mirjalili et al., 2014)
2. WOA — Whale Optimization Algorithm (Mirjalili & Lewis, 2016)
3. ALO — Ant Lion Optimizer (Mirjalili, 2015)
4. MFO — Moth-Flame Optimization (Mirjalili, 2015)
5. GOA — Grasshopper Optimization Algorithm (Saremi et al., 2017)
6. MVO — Multi-Verse Optimizer (Mirjalili et al., 2016)
7. SCA — Sine Cosine Algorithm (Mirjalili, 2016)
8. AOA — Arithmetic Optimization Algorithm (Abualigah et al., 2021)
9. MGO — Mountain Gazelle Optimizer (Abdollahzadeh et al., 2022)
10. GMO — Geometric Mean Optimizer (Mirrashid & Naderpour, 2023)

---

📐 Decision & Evaluation Modes

1. Metric Champions

Identifies individual objective leaders: Highest Top-1 Accuracy, Lowest Latency (ms), Smallest Footprint (MB), and Lowest Energy (J).

2. Weighted Sum Model (WSM) Composite Scoring

Normalizes all 4 objectives to $[0, 1]$ (inverting latency, size, and energy for minimization) and computes composite score: $$\text{Score} = \left( w_{\text{acc}} \cdot \tilde{A} + w_{\text{lat}} \cdot \tilde{L} + w_{\text{size}} \cdot \tilde{S} + w_{\text{energy}} \cdot \tilde{E} \right) \times 100$$

3. Pareto Non-Dominance Frontier

Identifies all non-dominated candidate models. A solution $\mathbf{A}$ dominates $\mathbf{B}$ if $\mathbf{A}$ is superior in at least one objective and no worse in any objective.

---

👨‍💻 Author & Maintainer

Created and maintained by Umesh (@UmeshCode1).