Bayesian
A Bayesian Optimization (BO) solver for stochastic simulation optimization. BO is a sequential, model-based method for expensive, noisy objectives: it fits a probabilistic surrogate to the observed data and, each iteration, optimizes a cheap acquisition function over the surrogate (no simulation) to choose the single most promising point to evaluate next.
The loop:
initializeIterations evaluates a space-filling initial design (one batch oracle call), fits the surrogate (and its hyperparameters), and records the incumbent.
Each mainIteration: (optionally) refit hyperparameters, refit the surrogate to the archive, compute the incumbent, maximize the acquisition over the surrogate, and evaluate the chosen point — the only oracle call of the iteration.
Per-point observation noise is taken from each Solution's estimated response (sample variance divided by replication count), which is exactly what Gaussian-process regression needs for the stochastic setting. The best solution found is tracked automatically by the base class through the currentSolution setter. All randomness is drawn through the solver's single random number stream (rnStream).
Parameters
the problem being solved
the evaluator responsible for assessing the quality of solutions
the random number stream number; 0 (the default) means the next available stream
the provider of random number streams; defaults to a fresh RNStreamProvider
the surrogate model; defaults to a GaussianProcessModel
the acquisition function; defaults to ExpectedImprovement
the acquisition optimizer; defaults to SampledAcquisitionOptimizer
the surrogate hyperparameter fitter; defaults to FixedHyperparameters
the initial design strategy; defaults to LatinHyperCubeDesign
the incumbent rule; defaults to BestPosteriorMeanIncumbent
the number of initial design points
the maximum number of BO iterations (after the initial design)
strategy to determine the number of replications per evaluation
an optional name for the solver
Constructors
Constructs a Bayesian optimization solver using a fixed number of replications per evaluation.
Properties
The acquisition function. Cannot be changed while the solver is running.
The acquisition optimizer. Cannot be changed while the solver is running.
Structured, ordered snapshot of the solver's configuration — the same fields the toString implementations expose, in a form that downstream consumers can iterate (reporting, machine-readable summary artifacts, "compare two runs" tooling).
The current incumbent value according to the configured incumbentRule.
The surrogate hyperparameter fitter. Cannot be changed while the solver is running.
The incumbent rule. Cannot be changed while the solver is running.
The initial design strategy. Cannot be changed while the solver is running.
The number of initial design points. Must be >= 2 (recommended >= problem dimension + 1).
An optional cap on the surrogate's training-set (archive) size, bounding the GP's O(n^3) cost; when set, the best solutions are retained. Null (the default) means no cap. Must be
A read-only view of the observed solutions (the surrogate's training set).
How often (in iterations) the surrogate hyperparameters are refit. Must be >= 1.
Used to detect no-improvement convergence.
The surrogate model. Cannot be changed while the solver is running.
Functions
A hook for subclasses to inject their specific internal state metrics into the snapshot without having to override the entire makeSolverStateSnapshot method.
Subclasses may implement this function to prepare the solver before running the first iteration. Generally, it is sufficient to just implement the startingPoint() function.
Subclasses should implement this function to determine if the solver should continue running iterations. This will likely include some implementation of stopping criteria. This function should implement stopping criteria based on the quality of the solution. The number of iterations, compared to the maximum number of iterations, is automatically checked after each step in the iterative process. Unless overridden, this function returns false by default, which indicates that the solution quality criteria have not been satisfied. This will cause the solver to iterate through all iterations of the solution process up to the maximum number of iterations. Alternatively, the user can specify an instance of the SolutionQualityEvaluatorIfc interface to determine if the solution quality has been reached.
This function should contain the logic that iteratively executes until the maximum number of iterations is reached or until the stopping criteria is met. The base implementation calls nextPoint() to determine the next point to evaluate, requests an evaluation of the point, and then updates the current solution if the resulting solution is better than the current solution. Generally, implementing startingPoint() and nextPoint() should be adequate. The property iterationCounter represents the current iteration within the mainIteration() function. That is, the value of iterationCounter is incremented prior to the execution of the mainIteration() function.