Hyperparameter Optimization Bayesian

Hyperparameter Optimization (Bayesian, Optuna, Population-Based Training) is the systematic process of selecting optimal training configurations—learning rates, batch sizes, architectures, regularization strengths—that maximize model performance — replacing manual trial-and-error tuning with principled search algorithms that efficiently explore high-dimensional configuration spaces.

The Hyperparameter Challenge

Neural network performance is highly sensitive to hyperparameter choices: a 2x change in learning rate can mean the difference between convergence and divergence; batch size affects generalization; weight decay interacts non-linearly with learning rate and architecture. Manual tuning is time-consuming and biased by practitioner experience. The search space grows combinatorially—10 hyperparameters with 10 values each yields 10 billion combinations, making exhaustive search impossible.

Grid Search and Random Search

• Grid search: Evaluates all combinations of discrete hyperparameter values; scales exponentially O(k^d) where k is values per dimension and d is number of hyperparameters
• Random search (Bergstra and Bengio, 2012): Randomly samples configurations from specified distributions; provably more efficient than grid search when some hyperparameters matter more than others
• Why random beats grid: Grid search wastes evaluations exploring irrelevant hyperparameter dimensions uniformly; random search allocates more unique values to each dimension
• Practical recommendation: Random search with 60 trials covers the space well enough for many problems; serves as baseline for more sophisticated methods

Bayesian Optimization

• Surrogate model: Builds a probabilistic model (Gaussian Process, Tree-Parzen Estimator, or Random Forest) of the objective function from evaluated configurations
• Acquisition function: Balances exploration (uncertain regions) and exploitation (promising regions)—Expected Improvement (EI), Upper Confidence Bound (UCB), or Knowledge Gradient
• Sequential refinement: Each trial's result updates the surrogate model, and the next configuration is chosen to maximize the acquisition function
• Gaussian Process BO: Models the objective as a GP with RBF kernel; provides uncertainty estimates but scales poorly beyond ~20 dimensions and ~1000 evaluations
• Tree-Parzen Estimator (TPE): Models the distribution of good and bad configurations separately using kernel density estimation; handles conditional and hierarchical hyperparameters naturally; default algorithm in Optuna and HyperOpt

Optuna Framework

• Define-by-run API: Hyperparameter search spaces are defined within the objective function using trial.suggest_* methods, enabling dynamic and conditional parameters
• Pruning (early stopping): MedianPruner and HyperbandPruner terminate unpromising trials early based on intermediate results, saving 2-5x compute
• Multi-objective optimization: Simultaneously optimizes accuracy and latency/model size using Pareto-optimal trial selection (NSGA-II)
• Distributed search: Scales across multiple workers with shared storage backend (MySQL, PostgreSQL, Redis)
• Visualization: Built-in plotting for optimization history, parameter importance, parallel coordinate plots, and contour maps
• Integration: Direct support for PyTorch Lightning, Keras, XGBoost, and scikit-learn through callback-based pruning

Population-Based Training (PBT)

• Evolutionary approach: Maintains a population of models training in parallel, each with different hyperparameters
• Exploit and explore: Periodically, underperforming members copy weights from top performers (exploit) and perturb hyperparameters (explore)
• Online schedule discovery: PBT implicitly learns hyperparameter schedules (e.g., learning rate warmup then decay) rather than fixed values—discovering that optimal hyperparameters change during training
• DeepMind results: PBT discovered training schedules for transformers, GANs, and RL agents that outperform manually designed schedules
• Communication overhead: Requires shared filesystem or network storage for model checkpoints; population size of 20-50 is typical

Advanced Methods and Practical Guidance

• BOHB (Bayesian Optimization HyperBand): Combines Bayesian optimization (TPE) with Hyperband's adaptive resource allocation for efficient multi-fidelity search
• Multi-fidelity optimization: Evaluate configurations cheaply first (few epochs, subset of data, smaller model) and allocate full resources only to promising candidates
• Transfer learning for HPO: Warm-start optimization using results from related tasks or datasets, reducing required evaluations by 50-80%
• Learning rate range test: Smith's learning rate finder sweeps learning rate from small to large in a single epoch, identifying optimal range without full HPO
• Hyperparameter importance: fANOVA (functional ANOVA) decomposes objective variance to identify which hyperparameters matter most, focusing search on high-impact dimensions

Hyperparameter optimization has evolved from ad-hoc manual tuning to a principled engineering practice, with frameworks like Optuna and methods like PBT enabling practitioners to systematically discover training configurations that unlock the full potential of their neural network architectures.