github.com/thalesfsp/ho

ho (Hyperparameter Optimizer)

A powerful Go library for automated hyperparameter optimization using Bayesian optimization with Gaussian Processes. The package provides efficient, thread-safe optimization capabilities for tuning system parameters with minimal manual intervention.

Features

• Bayesian Optimization: Uses Gaussian Process regression to efficiently explore parameter spaces
• Thread-safe Implementation: All components are designed for concurrent optimization runs
• Multiple Acquisition Functions: Various strategies for parameter space exploration:
  • Upper Confidence Bound (UCB)
  • Probability of Improvement (PI)
  • Expected Improvement (EI)
  • Thompson Sampling
• Generic Implementation: Works with both integer and floating-point parameters
• Progress Monitoring: Real-time updates on optimization progress via channels
• Flexible Configuration: Highly customizable optimization process
• Automatic Parameter Tuning: Learns from previous evaluations to suggest better parameters
• Robust Error Handling: Comprehensive error handling for benchmark functions

Install

go get github.com/yourusername/ho

Acquisition Functions

The library provides four acquisition functions for different optimization strategies:

1. Upper Confidence Bound (UCB)

• Balances exploration and exploitation
• Controlled by Beta parameter (higher = more exploration)
• Default choice, works well in most cases

config := DefaultConfig()  // Uses UCB by default
config.AcqParams.Beta = 2.0  // Adjust exploration-exploitation trade-off

2. Probability of Improvement (PI)

• Conservative exploration strategy
• Focuses on small, reliable improvements
• Good for noise-sensitive applications

config := DefaultConfig()
config.AcquisitionFunc = ProbabilityOfImprovement
config.AcqParams.Xi = 0.01  // Minimum improvement threshold

3. Expected Improvement (EI)

• Balances improvement probability and magnitude
• Most commonly used in practice
• Good for general optimization tasks

config := DefaultConfig()
config.AcquisitionFunc = ExpectedImprovement
config.AcqParams.Xi = 0.01  // Minimum improvement threshold

4. Thompson Sampling

• Simple but effective random sampling approach
• Great for parallel optimization
• No parameter tuning required

config := DefaultConfig()
config.AcquisitionFunc = ThompsonSampling
config.AcqParams.RandomState = rand.New(rand.NewSource(time.Now().UnixNano()))

Configuration

The OptimizationConfig struct allows customization of the optimization process:

type OptimizationConfig struct {
    Iterations      int              // Number of optimization steps
    InitialSamples  int              // Initial random samples
    NumCandidates   int              // Candidates per iteration
    AcquisitionFunc AcquisitionFunc  // Strategy for point selection
    AcqParams       AcquisitionParams // Parameters for acquisition function
    ProgressChan    chan<- ProgressUpdate // For progress monitoring
}

Recommended settings:

• Iterations: 20-200 (more = better results but longer runtime)
• InitialSamples: 5-20 (more = better initial model)
• NumCandidates: 50-500 (more = better search but slower iterations)

Thread Safety

All components are designed to be thread-safe:

• Safe for concurrent optimization runs with different configs
• Gaussian Process model uses RWMutex for thread-safe updates
• Progress channel updates are properly synchronized
• Random number generation is thread-safe

Contributing

• Fork the repository
• Clone your fork
• Create a feature branch
• Make your changes
• Run tests
• Create a pull request