Bayesian Optimization (貝式優化)

Bayesian Optimization is a probabilistic model-based optimization technique used for finding the global optimum of objective functions that are expensive to evaluate, non-convex, and lack an analytical expression. It is widely applied in domains like hyperparameter tuning, design optimization, and experiment control, where evaluating the objective function is costly in terms of time, resources, or computation.

How Bayesian Optimization Works

1. Surrogate Model:
   • Bayesian Optimization uses a surrogate probabilistic model, commonly a Gaussian Process (GP), to approximate the objective function based on the observed data (evaluations of the function). The surrogate model provides a probability distribution over possible function values, giving both a mean and uncertainty estimate.
2. Acquisition Function:
   • An acquisition function determines the next point to evaluate by balancing exploration (searching areas with high uncertainty) and exploitation (searching areas likely to improve the objective value). Common acquisition functions include:
     • Expected Improvement (EI): Chooses points that are expected to improve the objective function.
     • Probability of Improvement (PI): Selects points with a high probability of surpassing the current best value.
     • Upper Confidence Bound (UCB): Prioritizes points with high predicted mean and uncertainty.
3. Iterative Process:
   • Bayesian Optimization alternates between:
     1. Selecting a point to evaluate based on the acquisition function.
     2. Updating the surrogate model with the new data point.
     3. Repeating until a stopping criterion (e.g., maximum iterations, budget, or desired accuracy) is met.

Key Features of Bayesian Optimization

1. Sample Efficiency:
   • By leveraging prior evaluations to guide future evaluations, it minimizes the number of objective function evaluations required to find the optimum.
2. Uncertainty Quantification:
   • The surrogate model provides confidence intervals for predictions, allowing informed exploration of regions with high uncertainty.
3. Applicability to Black-Box Functions:
   • It works well for functions that are noisy, discontinuous, or do not have an analytical form.
4. High Computational Cost Justification:
   • Ideal for scenarios where each function evaluation is expensive, such as simulations, experiments, or complex model training.

Applications of Bayesian Optimization

1. Hyperparameter Tuning:
   • Used in machine learning to optimize hyperparameters of algorithms like neural networks, SVMs, and random forests. Frameworks like Optuna, Spearmint, and GPyOpt implement Bayesian Optimization for this purpose.
2. Experiment Design:
   • Helps optimize the configuration of physical experiments (e.g., in material science or chemistry) to maximize results with minimal trials.
3. A/B Testing:
   • Guides the selection of experimental variants in marketing or product design to optimize outcomes efficiently.
4. Robotics and Control:
   • Used for tuning control parameters in robots or optimizing trajectories.
5. Portfolio Optimization:
   • Applied in finance to optimize investment portfolios by balancing risk and return.
6. Design of Neural Architectures:
   • Guides neural architecture search (NAS) to find efficient deep learning models.
7. Reinforcement Learning:
   • Optimizes policies in reinforcement learning tasks, especially in scenarios with expensive simulations.
8. Chemical and Drug Discovery:
   • Finds optimal molecular configurations or reaction conditions for desired properties.

Advantages of Bayesian Optimization

1. Efficient Sampling:
   • Reduces the need for exhaustive search by intelligently selecting points to evaluate.
2. Flexibility:
   • Applicable to high-dimensional, noisy, and black-box optimization problems.
3. Informed Decision-Making:
   • The probabilistic nature of the surrogate model allows uncertainty-driven exploration.
4. Global Optimization:
   • Effective for finding global optima in functions with multiple local minima.