Functions

bo!(problem::BossProblem{Function}; kwargs...)
x = bo!(problem::BossProblem{Missing}; kwargs...)

Run the Bayesian optimization procedure to solve the given optimization problem or give a recommendation for the next evaluation point if problem.f == missing.

Arguments

• problem::BossProblem: Defines the optimization problem.

Keywords

• model_fitter::ModelFitter: Defines the algorithm used to estimate model parameters.
• acq_maximizer::AcquisitionMaximizer: Defines the algorithm used to maximize the acquisition function.
• term_cond::TermCond: Defines the termination condition.
• options::BossOptions: Defines miscellaneous settings.

References

BossProblem, ModelFitter, AcquisitionMaximizer, TermCond, BossOptions

Examples

See 'https://soldasim.github.io/BOSS.jl/stable/example/' for example usage.

estimate_parameters!(::BossProblem, ::ModelFitter)

Estimate the model parameters & hyperparameters using the given model_fitter algorithm.

Keywords

• options::BossOptions: Defines miscellaneous settings.

x = maximize_acquisition(::BossProblem, ::AcquisitionMaximizer)

Maximize the given acquisition function via the given acq_maximizer algorithm to find the optimal next evaluation point(s).

Keywords

• options::BossOptions: Defines miscellaneous settings.

eval_objective!(::BossProblem, x::AbstractVector{<:Real})

Evaluate the objective function and update the data.

Keywords

• options::BossOptions: Defines miscellaneous settings.

update_parameters!(problem::BossProblem, params::ModelParams)

Update the model parameters.

augment_dataset!(::BossProblem, x::AbstractVector{<:Real}, y, others...)
augment_dataset!(::BossProblem, x::AbstractVector{<:Real}, outputs::Any)

Add a new data point to the dataset. The optional others arguments are passed to the data-specific augment_dataset method.

construct_acquisition(::AcquisitionFunction, ::BossProblem, ::BossOptions) -> (x -> ::Real)

Construct the given AcquisitionFunction for the given BossProblem.

The returned function acq(x::AbstractVector{<:Real}) = val::Real is maximized to select the next evaluation function of blackbox function in each iteration.

This method must be implemented for all subtypes of AcquisitionFunction.

model_posterior(::BossProblem) -> ::Union{<:ModelPosterior, <:AbstractVector{<:ModelPosterior}}
model_posterior(::SurrogateModel, ::ModelParams, ::ExperimentData) -> ::ModelPosterior
model_posterior(::SurrogateModel, ::AbstractVector{<:ModelParams}, ::ExperimentData) -> ::AbstractVector{<:ModelPosterior}

Return an instance of ModelPosterior allowing to evaluate the posterior predictive distribution, or a vector of ModelPosteriors in case of multiple sampled model parameters.

See ModelPosterior for the list of available methods.

model_posterior_slice(::BossProblem, slice::Int) -> ::Union{<:ModelPosteriorSlice, <:AbstractVector{<:ModelPosteriorSlice}}
model_posterior_slice(::SurrogateModel, ::ModelParams, ::ExperimentData, slice::Int) -> ::ModelPosteriorSlice
model_posterior_slice(::SurrogateModel, ::AbstractVector{<:ModelParams}, ::ExperimentData, slice::Int) -> ::AbstractVector{<:ModelPosteriorSlice}

Return an instance of ModelPosteriorSlice allowing to evaluate the posterior predictive distribution, or a vector of ModelPosteriorSlices in case of multiple sampled model parameters.

See ModelPosteriorSlice for the list of available methods.

predictive_kind(::Type{<:SurrogateModel}) -> ::PredictiveKind
predictive_kind(::SurrogateModel) -> ::PredictiveKind
predictive_kind(::AbstractModelPosterior) -> ::PredictiveKind

Trait indicating whether the model's posterior predictive distribution is genuinely Gaussian (GaussianPredictive(), the default — mean/var/mean_and_var are exact) or whether the model additionally provides a more precise weighted-sample representation via predictive_samples (SampledPredictive()).

Should be implemented only on the model's type, i.e. predictive_kind(::Type{<:CustomModel}) = SampledPredictive(). The instance- and posterior-level methods (defined generically here and in ModelPosterior/ModelPosteriorSlice) forward to this type-level method automatically and should not be overridden separately — this keeps the trait's answer for a model, its instances, and its posteriors always in sync by construction.

A model returning SampledPredictive() should implement predictive_samples on its ModelPosterior/ModelPosteriorSlice type(s). A model whose predictive is genuinely Gaussian (e.g. GaussianProcess) should leave this at the default GaussianPredictive() rather than implementing a redundant, strictly noisier sample-based representation.

A wrapper model that reuses another model's predictive distribution unchanged (e.g. TransformedModel) should carry that wrapped model's type as one of its own type parameters and dispatch predictive_kind directly off it — see TransformedModel for the pattern.

predictive_samples(::ModelPosteriorSlice, ::AbstractVector{<:Real}) -> ::Tuple{<:AbstractVector{<:Real}, <:AbstractVector{<:Real}}
predictive_samples(::ModelPosteriorSlice, ::AbstractMatrix{<:Real}) -> ::Tuple{<:AbstractMatrix{<:Real}, <:AbstractMatrix{<:Real}}
predictive_samples(::ModelPosterior, ::AbstractVector{<:Real}) -> ::Tuple{<:AbstractMatrix{<:Real}, <:AbstractMatrix{<:Real}}
predictive_samples(::ModelPosterior, ::AbstractMatrix{<:Real}) -> ::Tuple{<:AbstractArray{<:Real, 3}, <:AbstractArray{<:Real, 3}}

Return a weighted discretization {(yₖ, wₖ)} of the (possibly non-Gaussian) posterior predictive distribution at the given point(s), with sum(ws) == 1. "Samples" may be deterministic quadrature nodes, weighted MC/MCMC draws, or any other weighted-point representation.

Shapes

• ModelPosteriorSlice, point x: ys, ws are length-K vectors.
• ModelPosteriorSlice, points X ((x_dim, n)): Ys, Ws are (K, n).
• ModelPosterior, point x: Ys is (y_dim, K); Ws is (1, K)
• ModelPosterior, points X: Ys is (y_dim, K, n), Ws is (1, K, n).

Ws's shared weight across dimensions must hold by construction of the sampling scheme (e.g. shared quadrature nodes); it cannot be derived after the fact from differing per-dimension weights.

Only implemented by models with predictive_kind(...) isa SampledPredictive; assumed more precise than mean/var/mean_and_var. Calling this on a GaussianPredictive model throws MethodError.

Implement at exactly one level

Implement this only at ModelPosteriorSlice (independent per-dimension sampling, e.g. WarpedGaussianProcess) or only at ModelPosterior (genuine joint sampling) — never both, and never by bundling one from the other. Bundling independent per-dimension atoms under a shared index induces artificial dependency between dimensions that the model doesn't actually have, giving silently wrong results for any likelihood that isn't separable across dimensions. See posterior.jl for how DefaultModelPosterior/DefaultModelPosteriorSlice handle this.

See also: predictive_kind.

x_dim(::BossProblem) -> Int

Return the input dimension of the problem.

y_dim(::BossProblem) -> Int

Return the output dimension of the problem.

cons_dim(::BossProblem) -> Int
cons_dim(::Domain) -> Int

Return the output dimension of the constraint function on the input.

See Domain for more information.

data_count(::BossProblem) -> Int

Return the number of datapoints in the dataset.

is_consistent(::BossProblem) -> Bool

Return true iff the model parameters have been fitted using the current dataset.

get_fitness(::AcquisitionFunction) -> (y -> ::Real)

Return the fitness function if the given AcquisitionFunction defines it. Otherwise, throw MethodError.

get_params(::BossProblem) -> ::Union{::ModelParams, AbstractVector{<:ModelParams}}
get_params(::UniFittedParams) -> ::ModelParams
get_params(::MultiFittedParams) -> ::AbstractVector{<:ModelParams}

Return the fitted ModelParams or a vector of ModelParams samples.

result(problem) -> (x, y)

Return the best found point (x, y).

The provided BossProblem must contain an AcquisitionFunction, with the get_fitness function defined.

Returns the point (x, y) from the dataset of the given problem such that y satisfies the constraints and fitness(y) is maximized. Returns nothing if the dataset is empty or if no feasible point is present.

Does not check whether x belongs to the domain as no exterior points should be present in the dataset.

Return an Inverse Gamma distribution with approximately 0.99 probability mass between lb and ub.

TruncatedMvNormal(μ, Σ, lb, ub)

Defines the truncated multivariate normal distribution with mean μ, covariance matrix Σ, lower bounds lb, and upper bounds ub.