Swarm optimization, from a single algorithm to a coordinated fleet.
A .NET library of swarm and metaheuristic optimizers, with the tooling to run them, measure them, and watch them converge.
HyperSwarm is a collection of population-based optimizers for hard search problems: hyperparameter tuning, scheduling, routing, and continuous and multi-objective design spaces.
It ships the classic algorithms, hybrids that combine them, and a coordination layer that runs several swarms together and decides which ones deserve more budget. Characterization and visualization tools sit alongside the algorithms, so you can see why a run behaved the way it did.
Allocates compute budget dynamically to high-performing subswarms.
Offloads velocity updates, dominance checks, and hypervolume to Intel GPUs.
Powers telemetry-aware architecture exploration inside AutoMLApp.
From single-particle mathematical operators to self-organizing multi-swarm ensembles across hardware backends.
11 core swarm algorithms (PSO, Ant Colony, Bee Colony, Bat, Grey Wolf, Fish School, Firefly, Glowworm, Cuckoo, Krill Herd, Synchronizing Firefly), 14 hybrid metaheuristics, 7 advanced metaheuristics (including OfBatsAndBeacons, ILS, MAP-Elites), and 3 multi-objective engines.
MOPSO and NSGA-III-style PSO with real-time non-dominated sorting, Pareto frontier preservation, and GPU-accelerated Monte Carlo hypervolume measurement.
Run heterogeneous swarms as coordinated ensembles. Director components plan subswarm launches, assign authority dynamically, and route workload across distributed hardware fabrics.
Fuzzy inertia control, topology learning, landscape analysis, entropy-based diversity tracking, early stopping, and automatic perturbation restart strategies.
Indirect pheromone-like coordination through shared trace buffers, episodic policy memory, and warm cache stores so repeated searches don't start from scratch.
Deep runtime profiling, convergence trajectory monitors, spatial particle distribution visualizers, and detailed diagnostic logs explaining search decisions.
Native Level Zero SPIR-V / zebin kernel offloads for ACO transitions, CMA-ES covariance, DE mutation, dominance matrix checks, N-body repulsion, and pairwise distances.
Acts as the pluggable search engine behind AutoMLApp's hyperparameter exploration, balancing model architecture performance against telemetry cost signals.
Comprehensive visual inventory across 11 Core Swarms, 14 Hybrids, 7 Advanced Metaheuristics, and 3 Multi-Objective engines in SwarmAlgorithms.
35 verified optimizers • 11 Core Swarms • 14 Hybrids • 7 Advanced Metaheuristics • 3 Multi-Objective
Exact mathematical models inHyperSwarm.Optimizers — Click any tile to inspect parameters and GPU kernels Classical and inertia-weighted velocity update models with global/local best tracking.
Pheromone trail deposition and probabilistic path construction with GPU offload.
Employed, onlooker, and scout bee foraging phases for multimodal landscapes.
Echolocation, pulse emission frequency, and loudness attenuation tuning.
Alpha, Beta, Delta, and Omega hierarchical leadership hunting behavior.
Praying, swarming, chasing, and leaping behaviors for dense search.
Light intensity flashing, attractiveness decay, and brightness distance attraction.
Luciferin-induced variable neighborhood search for simultaneous multimodal peak capture.
Lévy flights and parasitic egg replacement mechanisms for scale-invariant exploration.
Movement induced by other individuals, foraging activity, and random directional diffusion.
Coupled phase-oscillator synchrony with spatial coordinate updates.
Velocity vectors combined with vector differences and binomial crossovers.
Wolf pack leadership guiding velocity acceleration bounds.
Continuous particle trajectory informed by discrete pheromone graphs.
Luminescence attraction applied to inertia-weighted particles.
Boltzmann probability criteria preventing local stagnation.
Tabu memory lists preventing revisited fitness basins.
Covariance matrix adaptation applied to sub-clusters of swarm particles.
Bee colony scout mutations replaced by differential vector shifts.
Chromosome crossover applied to high-pheromone tour solutions.
Bee colony foraging informing particle personal best memories.
Global DE exploration transitioning into local SQP gradient polishing.
Flashing attractiveness enhanced with heavy-tailed Lévy flight leaps.
Encircling pack hunts combined with spiral bubble-net foraging maneuvers.
Individual particles execute localized line-search polishing at each iteration.
Models the search interplay of four archetypal characters—Batman, Robin, Gordon, and Joker—following their classic tropes (vigilante tracking, loyal support, signal coordination, and chaotic disruption).
Perturbation and local neighborhood search cycles for combinatorial spaces.
Multi-dimensional Archive of Phenotypic Elites illuminating diverse behavioral niches.
Replica exchange across heated energy levels to traverse rugged objective barriers.
Greedy Randomized Adaptive Search Procedure for complex routing and scheduling.
Dual constraint relaxation for heavily bounded optimization constraints.
Penalty parameter updates enforcing hard non-linear equality/inequality boundaries.
External archive of Pareto-optimal non-dominated leaders with crowding distance.
Reference-point-based non-dominated particle sorting for 3+ objective tasks.
Pluggable dominance comparator with dynamic hypervolume contribution tracking.
Topology learning dynamically shifts particle neighborhood connectivity based on entropy measurements and landscape variance, preventing early collapse into deceptive local optima.
Every capability is tagged with a distinct GUID/ID in the system registry, with reflection discovering additional plugin features at runtime.
Rate of change, stagnation detectors, diversity metrics, momentum gauges, and trajectory audits.
Fuzzy logic inertia tuners, adaptive mutation rates, topology adaptation, and auto-temperature schedulers.
Subswarm partitioning, migration managers, island models, and inter-swarm genetic exchange.
Dynamic budget allocator, supervisor authority matrices, task routing, and convergence arbitration.
Pheromone grid diffusion, episodic history recall, replay buffers, and warm-start caches.
Global best, Ring, Von Neumann, Random Geometric, Scale-Free, Multi-Swarm, Dynamic, and Dimension-Adaptive.
Non-dominated ranking, crowding distance, IGD+, hypervolume calculation, and reference-point niching.
KD-Tree spatial clustering, Voronoi particle partitioning, and neighbor lookups.
HyperSwarm executes mathematical hotspots directly on Intel Xe / Arc / Data Center Max GPUs via LevelZero.NET C ABI shims.
| Offload Target (10 Classes) | Compiled GPU Kernel | Mathematical Operation |
|---|---|---|
| ACO Transitions | pairwise_distance, fitness_kernel | GPU-accelerated tour edge selection and pheromone matrix updates. |
| CMA-ES Covariance | exchange_weights, tile_placement | Parallel matrix inversion and multivariate normal perturbation sampling. |
| DE Mutation & Crossover | fitness_kernel, rastrigin_fitness | Vector difference and binomial crossover computed across SIMD workgroups. |
| Dominance Matrix Checks | dominance_matrix | O(N²) all-to-all Pareto non-domination comparison performed in a single kernel launch. |
| Firefly Flashing & Distance | firefly_attraction, pairwise_distance | Quadratic distance attenuation and attractiveness updates parallelized across GPU threads. |
| Hypervolume Calculation | monte_carlo_hypervolume | High-throughput Monte Carlo sampling for multi-objective frontier hypervolume estimation. |
| N-Body & Swarm Repulsion | nbody_repulsion | Inter-particle electrostatic collision avoidance and diversity dispersion forces. |
| PSO Velocity Update | pso_velocity | Direct SIMD velocity and position updates across millions of particles per millisecond. |
// Build Director Ensemble with GPU acceleration and adaptive diversity budget
var ensemble = new SwarmDirectorBuilder()
.WithSubSwarm("pso_exploratory", SwarmMode.PSO, size: 200, topology: Topology.ScaleFree)
.WithSubSwarm("cuckoo_levy", SwarmMode.CuckooSearch, size: 80, topology: Topology.Ring)
.WithSubSwarm("mopso_pareto", SwarmMode.MOPSO, size: 150, topology: Topology.Dynamic)
.WithAcceleration(options =>
{
options.UseLevelZeroBackend = true;
options.PreferredDevice = DeviceSelector.IntelDiscreteGpu;
options.OffloadTargets = OffloadTarget.VelocityUpdate | OffloadTarget.DominanceMatrix | OffloadTarget.MonteCarloHypervolume;
})
.WithSelfAdaptation(adapt =>
{
adapt.TrackDiversityEntropy = true;
adapt.DynamicRestartThreshold = 0.04;
})
.Build();
var report = await ensemble.OptimizeAsync(objectiveFunc, searchBounds, CancellationToken.None);
Console.WriteLine($"Global Best Fitness: {report.BestFitness} (Evaluated {report.TotalEvaluations} points across {report.SubSwarmCount} swarms)");The algorithms and coordination layer are fully implemented and verified against approximately 2,500 test methods across 191 test files. HyperSwarm provides the high-performance search backend powering AutoMLApp's hyperparameter tuning and runs in harmony with SwarmDev's intent pipelines.
Visual canvas for hyperparameter search and hardware deployment.
Deterministic prompt-to-plan engine with wavescan bridge.
SIMD in-memory property graph with FastMCP agent context.
Detailed command-line benchmark invocation and metrics docs.