The 96% Speedup That Stops at the Optimizer

TL;DR for operators

MOSHO reports a 96.4% reduction in median optimizer runtime against NSGA-II when searching for neural architectures that balance predictive accuracy and device-specific inference latency. Across 18 benchmark contexts at 15,000 architecture evaluations, MOSHO takes 0.64 seconds versus 17.72 seconds for NSGA-II. Its normalized hypervolume is 0.9987, slightly below NSGA-II’s 0.9993.

The result is principally about making the search procedure cheaper. MOSHO directly edits categorical architecture choices, adapts its search operators using feedback, and avoids repeatedly performing expensive archive pruning.

For edge-AI engineering teams, the potential benefit is lower overhead when constructing candidate accuracy–latency trade-offs. The evidence does not establish equivalent reductions in neural-network training cost, physical energy consumption, or deployment latency. MOSHO actually reports slightly higher optimizer energy consumption than NSGA-II.

A dramatic speedup with almost no quality advantage

An engineering team selecting an image-classification model for an edge device faces competing requirements. A highly accurate architecture may be too slow for the intended hardware, while a faster architecture may sacrifice predictive performance. The team needs a selection of credible trade-offs before committing to a deployment configuration.

Finding those alternatives consumes computation. Faster search could therefore shorten model-development workflows, provided the architectures discovered remain competitive.

The MOSHO study by Aouni and colleagues¹ reports a striking runtime advantage against the established NSGA-II optimizer. Yet the quality difference is considerably smaller, and it favors NSGA-II on one central metric.

Source: Sections 5.5–5.6; 15,000 architecture evaluations across 18 dataset–device contexts.

A Pareto front contains architectures for which improving one objective requires sacrificing another. Normalized hypervolume measures how much of the relevant objective space a discovered front covers relative to the benchmark’s exact Pareto front. Values approaching one indicate strong coverage.

MOSHO reaches nearly complete coverage, but it does not outperform NSGA-II on median normalized hypervolume.

The result directs attention toward the computational work performed inside the optimizer. Understanding the speed advantage requires examining how architectures are represented, modified, and retained.

Categorical search eliminates an artificial numerical geometry

The original benchmark represents each architecture as six categorical choices drawn from five possible operations, including convolutions, pooling, skip connections, and a zero operation.

These categories have no meaningful continuous ordering. Moving numerically between operation indices does not correspond to a gradual architectural modification.

Earlier shark-inspired optimization methods were constructed around continuous positions and movement equations. Applying them to categorical architecture search requires converting numerical outputs back into discrete choices, potentially introducing artificial neighborhood relationships.

MOSHO instead proposes architectures through six native discrete operators:

• Patrol changes multiple architecture genes for broader exploration.
• Scent copies operations probabilistically from archive members, other agents, or random choices.
• Circle refines one or two genes locally.
• Burst combines partial elite copying with mutation.
• Crossover recombines categorical architecture genes.
• Scout introduces uniformly sampled architectures.

The optimizer also maintains a Pareto archive of promising, non-dominated solutions. This archive guides subsequent proposals rather than serving only as a final result collection.

Search agents receive an internal energy state that increases after improvements and decreases after unsuccessful proposals. Depleted agents restart through archive-based recombination or fresh exploration. This energy variable controls search behavior; it is unrelated to electrical consumption.

The contribution is an architecture-search process designed around the actual structure of the candidate space, without imposing continuous numerical movement on unordered categories.

Archive maintenance is a measurable source of overhead

Multi-objective optimizers must preserve several competing solutions. Their external archives therefore require dominance checks and diversity management as candidates arrive.

MOSHO separates these operations.

Dominated solutions are removed immediately, but expensive crowding-distance truncation is deferred until the end of a population iteration. The procedure preserves extreme solutions representing important boundaries of the objective space.

In the authors’ archive-management complexity analysis, the difference is substantial:

Here, $N$ denotes population size and $A$ denotes archive size.

This is a theoretical comparison with MOWSO, MOSHO’s shark-inspired predecessor. The 96.4% observed runtime reduction, by contrast, uses NSGA-II as its comparator. The two findings support the broader efficiency argument but should not be treated as the same measurement.

For optimization-platform engineers, deferred archive maintenance is a potentially transferable implementation principle: retain necessary candidate-level checks while reducing how often expensive diversity calculations occur.

Adaptive scheduling contributes more than every individual restart mechanism

MOSHO also adjusts how frequently it applies its six operators.

Its scheduler uses two signals: the archive improvement rate over the preceding five iterations and the fraction of the evaluation budget already consumed.

These inputs shift relative operator weights. Broad Patrol exploration receives greater emphasis early, while Scent and Crossover gain weight later. Other operators respond to recent progress or stagnation.

The single-component ablations in Section 5.4 identify this dual-signal scheduler as the strongest individual contributor to reported Pareto-front quality. Removing it produces the largest individual deterioration in normalized hypervolume and IGD+, with Vargha–Delaney $A_{12}=1.0$ favoring MOSHO on both metrics.

Other components receive less decisive standalone support. Energy-regulated restarts and restart-source mixing show comparatively limited individual effects. The final configuration also excludes three development-stage mechanisms: an explicit acceptance policy, K-means-based archive leader selection, and a mono-objective fallback.

Grouped ablations show larger deterioration when the search system’s core mechanisms are removed together.

The evidence supports a specific combination of categorical search, archive feedback, adaptive scheduling, and maintenance policies. It does not show that every proposed feature independently creates a large improvement.

The extensions test scope, not production readiness

The primary comparison uses NAS-Bench-201 and HW-NAS-Bench lookup tables spanning three image-classification datasets and six hardware targets. Its original search space contains 15,625 architectures, and experiments run 30 independent trials per method and context.

This setup allows controlled comparisons against exact reference Pareto fronts without repeatedly training networks. The 15,000-evaluation budget is close to the size of the original search space, although evaluation calls are not necessarily unique architectures.

The paper additionally reports two exploratory extensions.

In a larger 10-edge space containing 9,765,625 possible architectures, MOSHO’s normalized hypervolume degrades by less than 2%. However, this experiment uses LightGBM-predicted accuracy and latency, not directly measured outcomes for those newly constructed architectures.

In the three-objective accuracy–latency–energy experiment, MOSHO reaches normalized hypervolume of 0.9658, versus 0.9350 for NSGA-III and 0.9070 for MOEA/D. It also reports better IGD+, spacing, and execution time in that comparison.

These results support further investigation of scalability and additional objectives. They do not demonstrate performance under arbitrary architecture families, physical hardware variability, or production deployment constraints.

The supplementary sensitivity analysis also reports separate runtime figures of 0.9779 and 1.0047 seconds at the nominal 15,000-evaluation, 50-agent setting. Those values belong to different sensitivity tables and should not be substituted for the main comparison’s 0.64-second median without reconciling their aggregation contexts.

Faster execution does not automatically mean lower energy consumption

One result sharply limits the economic interpretation of MOSHO’s speed advantage.

At 15,000 evaluations, the paper reports optimizer energy consumption of 0.00052 kWh for MOSHO, compared with 0.00048 kWh for NSGA-II and 0.00045 kWh for MOPSO.

Thus, the algorithm with substantially shorter reported runtime than NSGA-II does not report lower computational energy consumption in the same study.

More fundamentally, lookup-table optimizer runtime excludes the expense of obtaining benchmark measurements, training candidate architectures, and operating the resulting networks.

For an edge-AI team estimating return on investment, these costs belong in separate accounting categories.

Cognaptus’ operational inference is that MOSHO deserves consideration when architecture evaluations are already inexpensive, objective measurements are reliable, and optimizer overhead materially affects search throughput. Under those conditions, reduced archive-management work and adaptive categorical exploration could accelerate candidate selection.

Where training dominates search expenditure, the reported runtime advantage may contribute little to total project cost. Where deployment decisions depend on memory footprint, power ceilings, device temperature, or workload-specific latency variability, those constraints require direct evaluation.

Before using MOSHO’s benchmark speedup in a production business case, engineering teams would need training-based comparisons, physical target-device measurements, and a cost breakdown separating optimizer execution from architecture evaluation and deployment.

The optimization lesson extends beyond shark-inspired search

MOSHO’s strongest contribution is not evidence that shark-inspired algorithms generally outperform evolutionary optimization. NSGA-II remains slightly ahead on the principal normalized-hypervolume comparison, and other baselines lead on some individual metrics.

The more defensible finding is that a carefully structured discrete optimizer can achieve competitive Pareto-front coverage while substantially reducing measured search overhead in controlled tabular benchmarks.

Its categorical operators address a representation mismatch. Its deferred truncation reduces archive-maintenance work. Its adaptive scheduler receives strong individual ablation support.

Together, these mechanisms give optimization-platform designers concrete techniques to investigate.

The production opportunity is faster construction of candidate architecture trade-offs. Whether that becomes cheaper model development or better deployed edge-AI systems remains an empirical question beyond the reported benchmarks.

Cognaptus: Automate the Present, Incubate the Future.