Supported Samplers

Tunny can utilize a number of sampling techniques.

The table below summarizes the optimizations supported by each sampler. Note that Tunny's UI will automatically display the available methods depending on the problem.

Name	Single-Objective	Multi-Objective	Constraints	Human-in-the-loop
AUTO Sampler	✓	✓	✓
TPE	✓	✓	✓	✓
cTPE	✓		✓
GP-Optuna	✓	✓	✓
GP-BoTorch	✓	✓	✓
GP-Preferential				✓
HEBO	✓
NSGA-II	✓	✓	✓
NSGA-III	✓	✓	✓
MOEA/D		✓
DE	✓
CMA-ES	✓
MO-CMA-ES		✓
INGO	✓
Random	✓	✓
QMC	✓	✓
BruteForce	✓	✓

The specific characteristics of each sampling technique are as follows

Name	Type	Note
AUTO Sampler	Auto	Automatically selects the optimal sampler based on problem characteristics (objectives, constraints, variable types). Recommended for beginners or when unsure which sampler to use.
TPE	Bayesian Optimization	Tree-structured Parzen Estimator. One of the most versatile methods alongside NSGA-II. Builds probabilistic models to efficiently explore search spaces. Excellent for expensive function evaluations and high-dimensional problems with mixed parameter types.
cTPE	Bayesian Optimization	Constrained TPE with enhanced constraint handling. Models constraint violations explicitly alongside objectives. Use when standard TPE struggles with constraint satisfaction or when strict constraint requirements exist.
GP-Optuna	Bayesian Optimization	Gaussian Process with Optuna's implementation. Faster than GP-BoTorch while maintaining good optimization quality. Provides uncertainty estimates and works well for smooth, continuous objective functions with moderate dimensionality.
GP-BoTorch	Bayesian Optimization	Gaussian Process using Facebook's BoTorch library. Highly flexible with advanced acquisition functions and state-of-the-art algorithms. Slower but offers superior optimization quality. Best when optimization quality is more important than speed.
GP-Preferential	Bayesian Optimization	Designed exclusively for Human-in-the-loop optimization. Learns from user preferences and pairwise comparisons instead of numerical objectives. Ideal for design optimization based on aesthetic preferences or subjective quality assessment.
HEBO	Bayesian Optimization	Heteroscedastic Evolutionary Bayesian Optimization. State-of-the-art algorithm that excels at highly nonlinear and multimodal problems. Combines evolutionary strategies with Bayesian optimization for faster convergence on challenging landscapes.
NSGA-II	Evolutionary Algorithm	Non-dominated Sorting Genetic Algorithm II. Widely used versatile method (also used in Wallacei). Uses non-dominated sorting and crowding distance to maintain diversity. Recommended for 2-3 objective problems requiring diverse Pareto fronts.
NSGA-III	Evolutionary Algorithm	Enhanced NSGA-II for many-objective optimization (3+ objectives). Uses reference points to maintain diversity in high-dimensional objective spaces. Better convergence and diversity for 4 or more objectives compared to NSGA-II.
MOEA/D	Evolutionary Algorithm	Multi-Objective EA based on Decomposition. Decomposes multi-objective problems into single-objective subproblems using scalarization. Efficient for 3- objectives with good computational performance and well-distributed Pareto fronts.
DE	Evolutionary Algorithm	Differential Evolution. Robust global optimization using vector differences. Effective for continuous, non-differentiable functions with many local optima. Simple yet powerful for rugged fitness landscapes.
CMA-ES	Evolutionary Strategy	Covariance Matrix Adaptation Evolution Strategy. One of the most powerful single-objective continuous optimizers. Adapts covariance matrix to learn problem structure. Very fast convergence on smooth problems with self-adaptive search distribution.
MO-CMA-ES	Evolutionary Strategy	Multi-objective extension of CMA-ES. Maintains multiple search distributions for different Pareto front regions. Fast convergence for 2-3 objectives on smooth landscapes. Efficient use of function evaluations.
INGO	Evolutionary Strategy	Implicit Natural Gradient Optimizer. Uses information geometry and natural gradient methods for efficient single-objective optimization. Fast convergence on well-structured problems with limited function evaluations.
Random	Sampling	Pure random sampling with uniform distribution. No optimization strategy. Useful for baseline performance, initial exploration, generating diverse populations, and testing purposes. Provides unbiased exploration of search space.
QMC	Sampling	Quasi-Monte Carlo using low-discrepancy sequences (e.g., Sobol). Better space coverage than random sampling with more even distribution. Reduced clustering, more efficient exploration. Effective for design of experiments and sensitivity analysis.
BruteForce	Sampling	Exhaustively evaluates all discrete variable combinations. Guarantees finding global optimum for discrete problems but only feasible for small search spaces. Computational cost grows exponentially. Use with caution.