HyperSwarm Ecosystem 34 Optimizers • 110 Test Functions Real-Time Telemetry & HUD

SwarmBench

Measure how swarms really behave, and watch them work.

HyperSwarm's benchmarking and visualization tools run 34 optimization algorithms against 110 test functions, record every run, and let you watch or replay each one agent by agent. Benchmark any algorithm on any problem with standardized seeds, eliminate underperforming configs in multi-stage tournaments, and inspect live swarm dispersion on desktop HUD canvases.

Explore 110 Functions CLI Commands Swarm Visualizer HUD HyperSwarm Product

34 Optimization Algorithms Classic swarms, search methods, 14 hybrids and multi-objective engines

110 Registered Test Functions 9 problem families: unimodal, multimodal, constrained, combinatorial, etc.

2,500+ Automated Tests Full unit test suite validating mathematical correctness and invariants

7 Run Lifecycle Stages From topology scan to convergence events and HyperCache writeback

100% Deterministic & Seeded Exact seed replay, dimension binning, and self-validating characterization

🌐

Broad by default

Classic swarms, search methods, 14 hybrids and two multi-objective algorithms meet unimodal, multimodal, constrained, multi-objective, combinatorial, dynamic, niching, noisy and high-dimensional problems.

🎯

Repeatable

Every run takes a seed, the larger problem variants use fixed seeds, and tuned settings are cached per function and dimension bin.

👁️

Visible

A live and replay visualizer shows agents, subswarms, topology regions and feature planes as a run unfolds.

🛡️

Checked

The algorithm library alone carries more than 2,500 automated tests, and characterization profiles validate themselves before they are written.

CLI • SwarmBenchmarkRunner

Benchmark Any Algorithm on Any Problem

One command-line runner, SwarmBenchmarkRunner, covers single runs, streamed telemetry, tournaments, tuning sweeps and cache maintenance. Every run takes the same core options: --algorithm, --function, --dimensions (default 10), --iterations (default 500), --swarm-size (default 50) and a random --seed (default 42).

Command What it does Example Syntax

list-functions, list-algorithms Discovery

Print the test-function and algorithm catalogues as JSON

SwarmBenchmarkRunner list-functions --json

run Execution

Run one benchmark with the chosen algorithm and function

SwarmBenchmarkRunner run --algorithm pso --function rosenbrock --dimensions 10 --iterations 500 --seed 42

run-full Execution

Run with telemetry and optional feature layers

SwarmBenchmarkRunner run-full --algorithm gwo --function ackley --features topology,hotspots

run-stream Telemetry

Emit real-time JSON telemetry on standard output, with a configurable sample interval

SwarmBenchmarkRunner run-stream --algorithm fa --function rastrigin --sample-interval-ms 16

run-diagnostic Diagnostics

Write a per-agent, per-iteration JSONL trace for replay and analysis

SwarmBenchmarkRunner run-diagnostic --algorithm mopso --function zdt1 --out trace.jsonl

run-game Tournament

Run a multi-stage elimination tournament across the test-function gamut, with survivor and election fractions, replicas and optional subswarm splitting

SwarmBenchmarkRunner run-game --pool 50 --survivors 0.5 --split-subswarms --backend levelzero

run-convergence-benchmark Analysis

Compare three search strategies, GlobalGrid, Beeline and Wavefront-1Iter, under an evaluation budget

SwarmBenchmarkRunner run-convergence-benchmark --function sphere --budget 800

run-convergence-benchmark-matrix Analysis

Sweep several budgets (default 200, 400, 800, 1600) and print a compact winner table per function

SwarmBenchmarkRunner run-convergence-benchmark-matrix --budgets 200,400,800,1600

run-hpo-sweep Tuning

Run an ASHA/BOHB-driven hyperparameter sweep with cached results

SwarmBenchmarkRunner run-hpo-sweep --algorithm pso_de --budget-type evaluations

train Training

Train swarm configurations per test function, filtered by category or dimension bin

SwarmBenchmarkRunner train --category multimodal --dimension-bin 10d

high-dim-bench Scaling

Run seeded high-dimensional benchmark expansions

SwarmBenchmarkRunner high-dim-bench --dimensions 100,200,500 --seed 42

hypercache-export, hypercache-validate, hypercache-validate-recall Cache

Export and validate the HyperCache of tuned configurations

SwarmBenchmarkRunner hypercache-validate --check-recall

Tournament Engine

Tournaments, not just single runs

The benchmark game puts a pool of agents (50 by default) through stages drawn from the test functions. Each stage elects agents for full trials, keeps a fraction as survivors, and can split the pool into subswarms once it converges together. Filter the stages by category, choose the algorithms in the pool, and set the hardware backend.

SwarmBenchmarkRunner run-game \
  --pool 50 \
  --survivors 0.5 \
  --split-subswarms \
  --backend levelzero

Tuned Cache

Tuned settings you can reuse

The HyperCache stores the best configurations found for each function and dimension bin, and the validate commands check that every cached configuration can still be reached across consecutive seed replays.

SwarmBenchmarkRunner hypercache-validate \
  --check-recall \
  --dimension-bins 10d,30d,100d

Self-Validating Reports

Characterization profiles

The SwarmCharacterization tool runs one algorithm on one function and dimension across several seeds (five by default). It self-validates the result before anything is written, then produces a machine-readable profile.json, a readable profile.md and a standalone LaTeX profile.tex per algorithm. Re-running a combination replaces only that entry, so profiles grow one function and dimension at a time.

SwarmCharacterization \
  --algorithm cma_es \
  --function rosenbrock \
  --seeds 5 \
  --emit json,md,tex

HUD • SwarmVisualizer.Host

Watch the Swarm Run

The visualizer turns any benchmark run into a scene you can watch live or replay later. It shows every agent, every subswarm and the landscape the swarm is reading.

Live Execution

Pipe a run straight in, or let the host launch and attach automatically.

SwarmBenchmarkRunner run-stream ... | SwarmVisualizer.Host stdin
# Or let the host launch the command:
SwarmVisualizer.Host capture <command>

Deterministic Replay

Plays back saved per-agent, per-iteration traces with variable playback speed.

SwarmVisualizer.Host replay trace.jsonl

Headless JSON Report

Turns a trace into an aggregated JSON run summary without rendering a window.

SwarmVisualizer.Host report trace.jsonl summary.json

What is on screen

The Windows desktop app draws a live scatter plot of the agents and time-series charts of the run's metrics, with run, source, iteration, progress, best fitness and agent count in view. It auto-attaches to local swarm telemetry, opens replay files, launches captured runs, and reloads saved runs from the persistence catalog. It sits on the same scene stream as the console host, so a run looks the same wherever you open it.

7 Lifecycle Stages of a Run

Run Started: Seed initialization, domain boundary allocation, and initial particle dispersion.

Topology Scan: Neighbourhood graph computation, communication adjacency matrices, and cluster seeds.

Iteration: Synchronous or asynchronous velocity, position, and fitness evaluations.

Improvement: Global or Pareto frontier advancement, pBest updates, and historical record tracking.

Convergence: Stagnation detection, entropy variance monitoring, and threshold verification.

Event: Subswarm split/merge, dynamic peak shift, obstacle encounter, or replica exchange.

Run Completed: Final metric telemetry aggregation, HyperCache writeback, and profile generation.

What a Frame Carries

Each frame delivers rich spatial and telemetry layers across the search landscape:

🐝 Agents

Position, velocity, fitness, personal best, age, neighbour count, agent type, subswarm and algorithm mode

👑 Subswarms

Parent and child links, depth, best fitness, stagnation count, active, grown or converged status

🌐 Topology Regions

Centre, estimated radius, average fitness, fitness contrast, deflection count, region type

📐 Feature Planes

Strength, confidence, urgency, activation height, suggested operation and preferred mode

🔥 Hotspots

Centroid, strength, average fitness, observation count

📊 Metrics

Named values with units, sampled as the run goes (convergence rate, diversity entropy, velocity variance)

Built to be Extended

Scenes come from pluggable sources (JSONL file, stdin, launched process, stream, persisted runs) and go to pluggable plugins (console summary, JSON summary, desktop window). Add your own by implementing:

public interface ISwarmSceneSource {
  IAsyncEnumerable<SwarmSceneFrame> ReadFramesAsync(CancellationToken ct);
}

public interface ISwarmScenePlugin {
  Task OnFrameReceivedAsync(SwarmSceneFrame frame, CancellationToken ct);
}

COMPLETE CATALOG (110 FUNCTIONS)

The 110 Registered Test Functions

The suite ships 110 registered test functions across nine problem families. Each reports recommended dimensions, default bounds and known optimum. Scaled variants share a name and differ by problem size so you can test high-dimensional scaling.

Function ID	Function Name	Category	Dimensions	Known Optimum	Formulation & Characterization Notes
`sphere`	Sphere	Unimodal	30D	`0`	Continuous, convex, isotropic bowl
`rosenbrock`	Rosenbrock	Unimodal	10D	`0`	Banana valley, non-separable gradient flow
`zakharov`	Zakharov	Unimodal	10D	`0`	Non-separable quadratic with power term
`dixonprice`	Dixon-Price	Unimodal	10D	`0`	Valley with quadratic index scaling
`booth`	Booth	Unimodal	2D	`0`	Plate-shaped quadratic benchmark
`bent_cigar`	Bent Cigar	Unimodal	10D	`0`	Extremely conditioned ill-scaled ridge
`powell`	Powell	Unimodal	24D	`0`	Quartic non-differentiable singularity
`hyper_ellipsoid`	Rotated Hyper-Ellipsoid	Unimodal	10D	`0`	Cumulative sum coordinate axes
`sum_of_diff_powers`	Sum of Different Powers	Unimodal	10D	`0`	Variable exponent steepness
`rastrigin`	Rastrigin	Multimodal	10D	`0`	Cosine lattice of local traps
`ackley`	Ackley	Multimodal	10D	`0`	Nearly flat outer region with deep hole
`griewank`	Griewank	Multimodal	10D	`0`	Product of cosine terms on parabola
`schwefel`	Schwefel	Multimodal	10D	`0`	Deceptive optimum at search space boundary
`levy`	Levy	Multimodal	10D	`0`	Sinusoidal modulation with sharp basin
`michalewicz`	Michalewicz	Multimodal	10D	`-9.66`	Steep valleys with factorial basin count
`beale`	Beale	Multimodal	2D	`0`	Sharp corner ridges
`goldstein_price`	Goldstein-Price	Multimodal	2D	`3`	Eighth-order polynomial landscape
`bukin_n6`	Bukin N.6	Multimodal	2D	`0`	Thin parabolic canyon with ripple waves
`eggholder`	Eggholder	Multimodal	2D	`-959.6407`	Extreme multimodality with chaotic basins
`cross_in_tray`	Cross-in-Tray	Multimodal	2D	`-2.0626`	Cross-shaped ridge with 4 global corners
`drop_wave`	Drop-Wave	Multimodal	2D	`-1`	Concentric circular ripple oscillations
`easom`	Easom	Multimodal	2D	`-1`	Microscopic basin surrounded by flat plane
`styblinski_tang`	Styblinski-Tang	Multimodal	10D	`-391.6617`	Quartic double-well per dimension
`schaffer_n2`	Schaffer N.2	Multimodal	2D	`0`	Circular ripple bands
`schaffer_n4`	Schaffer N.4	Multimodal	2D	`0.292579`	Asymmetric concentric ring barrier
`alpine`	Alpine N.1	Multimodal	10D	`0`	Absolute sine product peaks
`salomon`	Salomon	Multimodal	10D	`0`	Radial wave rings around origin
`hartmann3`	Hartmann 3D	Multimodal	3D	`-3.86278`	4 local minima in unit hypercube
`hartmann6`	Hartmann 6D	Multimodal	6D	`-3.32237`	6 local minima with non-linear exponential matrix
`happy_cat`	Happy Cat	Multimodal	10D	`0`	Curved valley with steep surrounding walls
`expanded_griewank_rosenbrock`	Expanded Griewank-Rosenbrock	Multimodal	10D	`0`	Nested Griewank cosine ripple inside Rosenbrock valley
`weierstrass`	Weierstrass	Multimodal	10D	`0`	Continuous everywhere, differentiable nowhere fractal
`g01`	G01	Constrained	13D	`-15`	CEC 2006, 9 linear inequality constraints
`g02`	G02	Constrained	20D	`-0.803619`	CEC 2006, 2 non-linear inequalities
`g03`	G03	Constrained	10D	`-1`	CEC 2006, 1 non-linear equality constraint
`g04`	G04	Constrained	5D	`-30665.539`	CEC 2006, 6 non-linear inequalities
`g05`	G05	Constrained	4D	`5126.4981`	CEC 2006, cubic equality constraints
`g06`	G06	Constrained	2D	`-6961.81388`	CEC 2006, cubic inequality constraint
`g07`	G07	Constrained	10D	`24.3062`	CEC 2006, 8 non-linear inequalities
`g08`	G08	Constrained	2D	`-0.095825`	CEC 2006, 2 non-linear inequality boundaries
`g09`	G09	Constrained	7D	`680.6301`	CEC 2006, 4 non-linear inequality constraints
`g10`	G10	Constrained	8D	`7049.3307`	CEC 2006, linear and cubic inequalities
`g11`	G11	Constrained	2D	`0.7499`	CEC 2006, quadratic equality constraint
`g12`	G12	Constrained	3D	`-1`	CEC 2006, 729 disjoint feasible sub-spheres
`g13`	G13	Constrained	5D	`0.0539498`	CEC 2006, 3 non-linear equality constraints
`g14`	G14	Constrained	10D	`-47.7649`	CEC 2006, 3 linear equality constraints
`g15`	G15	Constrained	3D	`961.715`	CEC 2006, 2 equality constraints
`g16`	G16	Constrained	5D	`-1.9051553`	CEC 2006, 38 linear/quadratic inequalities
`g17`	G17	Constrained	6D	`8853.5396`	CEC 2006, 4 non-linear equality constraints
`g18`	G18	Constrained	9D	`-0.8660254`	CEC 2006, 13 non-linear inequality constraints
`g19`	G19	Constrained	15D	`32.6555929`	CEC 2006, 5 non-linear equality constraints
`g21`	G21	Constrained	7D	`193.7245`	CEC 2006, linear and non-linear equalities
`g23`	G23	Constrained	9D	`-400.0551`	CEC 2006, 6 inequalities, 2 equalities
`g24`	G24	Constrained	2D	`-5.508`	CEC 2006, 2 non-linear inequalities
`pressure_vessel`	Pressure Vessel Design	Constrained	4D	`6059.714`	Engineering benchmark: Ts, Th, R, L design
`tension_compression_spring`	Tension/Compression Spring Design	Constrained	3D	`0.0126652`	Engineering benchmark: wire dia, mean coil dia, active coils
`welded_beam`	Welded Beam Design	Constrained	4D	`1.7249`	Engineering benchmark: weld thickness, length, bar depth, width
`zdt1`	ZDT1 2 Objectives	Multi-objective	30D	`Pareto Front [0, 1]`	Convex Pareto optimal front
`zdt2`	ZDT2 2 Objectives	Multi-objective	30D	`Pareto Front [0, 1]`	Non-convex Pareto optimal front
`zdt3`	ZDT3 2 Objectives	Multi-objective	30D	`Discontinuous Front`	5 disconnected Pareto optimal curves
`dtlz1`	DTLZ1 3 Objectives	Multi-objective	7D	`Linear Hyperplane`	11^k - 1 local Pareto fronts
`dtlz2`	DTLZ2 3 Objectives	Multi-objective	12D	`Spherical Sector`	Spherical Pareto frontier
`dtlz3`	DTLZ3 3 Objectives	Multi-objective	12D	`Spherical with Traps`	DTLZ2 front surrounded by dense local traps
`dtlz4`	DTLZ4 3 Objectives	Multi-objective	12D	`Biased Front`	Parametric density bias along coordinate axes
`dtlz5`	DTLZ5 3 Objectives	Multi-objective	12D	`Degenerate Curve`	Degenerate 1D curve in 3D objective space
`dtlz6`	DTLZ6 3 Objectives	Multi-objective	12D	`Extremely Degenerate`	Severe bias testing frontier dimensionality reduction
`dtlz7`	DTLZ7 3 Objectives	Multi-objective	22D	`Disconnected Regions`	2^(M-1) disconnected Pareto optimal regions
`dtlz1_m5`	DTLZ1 (many-objective) 5 Objectives	Multi-objective	9D	`5D Linear Hyperplane`	Many-objective 5D linear simplex
`dtlz2_m5`	DTLZ2 (many-objective) 5 Objectives	Multi-objective	14D	`5D Spherical Octant`	Many-objective 5D spherical sector
`dtlz7_m5`	DTLZ7 (many-objective) 5 Objectives	Multi-objective	24D	`5D Disconnected Fronts`	Many-objective 16 disconnected Pareto regions
`dtlz3_m8`	DTLZ3 (many-objective) 8 Objectives	Multi-objective	17D	`8D Spherical Hypersurface`	8-objective extreme multimodal traps
`dtlz4_m10`	DTLZ4 (many-objective) 10 Objectives	Multi-objective	19D	`10D Biased Manifold`	10-objective dense boundary bias
`dtlz5_m15`	DTLZ5 (many-objective) 15 Objectives	Multi-objective	24D	`15D Degenerate Manifold`	15-objective high-dimensional manifold test
`tsp_benchmark`	TSP Benchmark (Continuous)	Combinatorial	10D	`Global Shortest Tour`	Continuous relaxation encoding of TSP
`tsp_medium`	TSP (Medium)	Combinatorial	30D	`Fixed Seed Tour`	30-city seeded Euclidean permutation
`tsp_high`	TSP (High)	Combinatorial	100D	`Fixed Seed Tour`	100-city seeded Euclidean permutation
`tsp_very_high`	TSP (Very High)	Combinatorial	200D	`Fixed Seed Tour`	200-city large-scale route optimization
`qap_benchmark`	QAP Benchmark (Continuous)	Combinatorial	10D	`Optimal Facility Layout`	Quadratic Assignment Problem 10 facilities
`qap_medium`	QAP (Medium)	Combinatorial	30D	`Fixed Seed Assignment`	30 facilities, distance x flow tensor
`qap_high`	QAP (High)	Combinatorial	100D	`Fixed Seed Assignment`	100 facilities, large-scale assignment
`qap_very_high`	QAP (Very High)	Combinatorial	200D	`Fixed Seed Assignment`	200 facilities high-dimensional permutation
`knapsack`	0/1 Knapsack Problem	Combinatorial	20D	`Max Value Subset`	20 items, weight capacity constraint
`knapsack_medium`	Knapsack (Medium)	Combinatorial	30D	`Fixed Seed Max Value`	30 items with correlated weights
`knapsack_high`	Knapsack (High)	Combinatorial	100D	`Fixed Seed Max Value`	100 items strongly correlated profits
`knapsack_very_high`	Knapsack (Very High)	Combinatorial	200D	`Fixed Seed Max Value`	200 items high-dimensional knapsack
`moving_peaks`	Moving Peaks	Dynamic	5D	`Dynamic Peak`	5 moving cone peaks shifting in coordinate space
`moving_peaks_medium`	Moving Peaks (Medium)	Dynamic	30D	`Dynamic Peak`	8 peaks, changes every 3,000 evaluations
`moving_peaks_high`	Moving Peaks (High)	Dynamic	100D	`Dynamic Peak`	10 peaks, changes every 5,000 evaluations
`moving_peaks_very_high`	Moving Peaks (Very High)	Dynamic	200D	`Dynamic Peak`	12 peaks, changes every 7,000 evaluations
`peaks_5`	Moving Peaks Niching (5 peaks)	Dynamic	5D	`Dynamic Niche`	Changes every 5,000 evaluations
`peaks_10`	Moving Peaks Niching (10 peaks)	Dynamic	5D	`Dynamic Niche`	Changes every 5,000 evaluations
`peaks_20`	Moving Peaks Niching (20 peaks)	Dynamic	5D	`Dynamic Niche`	Changes every 5,000 evaluations
`df1`	DF1 Dynamic Function	Dynamic	10D	`Time-Varying Center`	CEC dynamic benchmark; the optimum position shifts over time
`equal_maxima`	Equal Maxima	Niching	1D	`1.0 (5 equal peaks)`	5 identical height sinusoidal peaks
`five_uneven_peak_trap`	Five Uneven Peak Trap	Niching	1D	`1.0 (Global)`	1 global peak and 4 deceptive sub-peaks
`uneven_decreasing_maxima`	Uneven Decreasing Maxima	Niching	1D	`1.0 (Exponential Decay)`	Exponentially attenuated peak heights
`himmelblau`	Himmelblau's Function	Niching	2D	`0 (4 global minima)`	4 identical global optima at symmetric coordinates
`six_hump_camel_back`	Six-Hump Camel Back	Niching	2D	`-1.0316 (2 global)`	2 global and 4 local minima
`shubert`	Shubert	Niching	2D	`-186.7309 (18 global)`	18 global and 742 local minima
`shubert_nd`	Shubert N-D	Niching	5D	`High-Dimensional Niches`	Exponentially exploding niche grid
`vincent`	Vincent	Niching	2D	`1.0 (36 global peaks)`	6^D = 36 global peaks with non-uniform spacing
`modified_rastrigin_niching`	Modified Rastrigin, all global optima	Niching	2D	`0 (Multiple Global)`	Modulated Rastrigin equalizing all local peaks
`noisy_sphere`	Noisy Sphere σ = 0.1	Noisy	30D	`0`	Gaussian noise sigma = 0.1 on evaluation
`noisy_sphere_high`	Noisy Sphere (High) σ = 0.1	Noisy	100D	`0`	100-dimensional noisy parabolic bowl
`noisy_sphere_very_high`	Noisy Sphere (Very High) σ = 0.1	Noisy	200D	`0`	200-dimensional noisy parabolic bowl
`noisy_rastrigin`	Noisy Rastrigin σ = 0.05	Noisy	10D	`0`	Multimodal traps + stochastic jitter sigma = 0.05
`noisy_rastrigin_high`	Noisy Rastrigin (High) σ = 0.1	Noisy	100D	`0`	100D multimodal traps with evaluation noise
`noisy_rastrigin_very_high`	Noisy Rastrigin (Very High) σ = 0.1	Noisy	200D	`0`	200D multimodal traps with evaluation noise
`highdim_sphere_low`	High-D Sphere (Low)	High-dimensional	10D	`0`	Seed expansion baseline (scales to 10,000D)
`highdim_rastrigin_medium`	High-D Rastrigin (Medium)	High-dimensional	30D	`0`	Multimodal seed expansion (scales to 5,000D)