Edit

Models

Models play an essential role in Bayesian Optimization (BO). A model is used as a surrogate function for the actual underlying black box function to be optimized. In BoTorch, a Model maps a set of design points to a posterior probability distribution of its output(s) over the design points.

In BO, the model used is traditionally a Gaussian Process (GP), in which case the posterior distribution is a multivariate normal. While BoTorch supports many GP models, BoTorch makes no assumption on the model being a GP or the posterior being multivariate normal. With the exception of some of the analytic acquisition functions in the botorch.acquisition.analytic module, BoTorch’s Monte Carlo-based acquisition functions are compatible with any model that conforms to the Model interface, whether user-implemented or provided.

Under the hood, BoTorch models are PyTorch Modules that implement the light-weight Model interface. When working with GPs, GPyTorchModel provides a base class for conveniently wrapping GPyTorch models.

Users can extend Model and GPyTorchModel to generate their own models. For more on implementing your own models, see Implementing Custom Models below.

Terminology

Multi-Output and Multi-Task

A Model (as in the BoTorch object) may have multiple outputs, multiple inputs, and may exploit correlation between different inputs. BoTorch uses the following terminology to distinguish these model types:

• Multi-Output Model: a Model with multiple outputs. Most BoTorch Models are multi-output.
• Multi-Task Model: a Model making use of a logical grouping of inputs/observations (as in the underlying process). For example, there could be multiple tasks where each task has a different fidelity. In a multi-task model, the relationship between different outputs is modeled, with a joint model across tasks.

Note the following:

• A multi-task (MT) model may or may not be a multi-output model. For example, if a multi-task model uses different tasks for modeling but only outputs predictions for one of those tasks, it is single-output.
• Conversely, a multi-output (MO) model may or may not be a multi-task model. For example, multi-output Models that model different outputs independently rather than building a joint model are not multi-task.
• If a model is both, we refer to it as a multi-task-multi-output (MTMO) model.

Noise: Homoskedastic, fixed, and heteroskedastic

Noise can be treated in several different ways:

• Homoskedastic: Noise is not provided as an input and is inferred, with a constant variance that does not depend on X. Many models, such as SingleTaskGP, take this approach. Use these models if you know that your observations are noisy, but not how noisy.

• Fixed: Noise is provided as an input and is not fit. In “fixed noise” models like FixedNoiseGP, noise cannot be predicted out-of-sample because it has not been modeled. Use these models if you have estimates of the noise in your observations (e.g. observations may be averages over individual samples in which case you would provide the mean as observation and the standard error of the mean as the noise estimate), or if you know your observations are noiseless (by passing a zero noise level).

• Heteroskedastic: Noise is provided as an input and is modeled to allow for predicting noise out-of-sample. Models like HeteroskedasticSingleTaskGP take this approach.

Standard BoTorch Models

BoTorch provides several GPyTorch models to cover most standard BO use cases:

Single-Task GPs

These models use the same training data for all outputs and assume conditional independence of the outputs given the input. If different training data is required for each output, use a ModelListGP instead.

• SingleTaskGP: a single-task exact GP that infers a homoskedastic noise level (no noise observations).
• FixedNoiseGP: a single-task exact GP that differs from SingleTaskGP in using fixed observation noise levels. It requires noise observations.
• HeteroskedasticSingleTaskGP: a single-task exact GP that differs from SingleTaskGP and FixedNoiseGP in that it models heteroskedastic noise using an additional internal GP model. It requires noise observations.
• MixedSingleTaskGP: a single-task exact GP that supports mixed search spaces, which combine discrete and continuous features.
• SaasFullyBayesianSingleTaskGP: a fully Bayesian single-task GP with the SAAS prior. This model is suitable for sample-efficient high-dimensional Bayesian optimization.

Model List of Single-Task GPs

• ModelListGP: A multi-output model in which outcomes are modeled independently, given a list of any type of single-task GP. This model should be used when the same training data is not used for all outputs.

Multi-Task GPs

• MultiTaskGP: a Hadamard multi-task, multi-output GP using an ICM kernel, inferring a homoskedastic noise level (does not require noise observations).
• FixedNoiseMultiTaskGP: a Hadamard multi-task, multi-output GP using an ICM kernel, with fixed observation noise levels (requires noise observations).
• KroneckerMultiTaskGP: A multi-task, multi-output GP using an ICM kernel, with Kronecker structure. Useful for multi-fidelity optimization.
• SaasFullyBayesianMultiTaskGP: a fully Bayesian multi-task GP using an ICM kernel. The data kernel uses the SAAS prior to model high-dimensional parameter spaces.

All of the above models use Matérn 5/2 kernels with Automatic Relevance Discovery (ARD), and have reasonable priors on hyperparameters that make them work well in settings where the input features are normalized to the unit cube and the observations are standardized (zero mean, unit variance).

Other useful models

• ModelList: a multi-output model container in which outcomes are modeled independently by individual Models (as in ModelListGP, but the component models do not all need to be GPyTorch models).
• SingleTaskMultiFidelityGP and FixedNoiseMultiFidelityGP: Models for multi-fidelity optimization. For more on Multi-Fidelity BO, see the tutorial.
• HigherOrderGP: A GP model with matrix-valued predictions, such as images or grids of images.
• PairwiseGP: A probit-likelihood GP that learns via pairwise comparison data, useful for preference learning.
• ApproximateGPyTorchModel: for efficient computation when data is large or responses are non-Gaussian.
• Deterministic models, such as AffineDeterministicModel, AffineFidelityCostModel, GenericDeterministicModel, and PosteriorMeanModel express known input-output relationships; they conform to the BoTorch Model API, so they can easily be used in conjunction with other BoTorch models. Deterministic models are useful for multi-objective optimization with known objective functions and for encoding cost functions for cost-aware acquisition.
• SingleTaskVariationalGP: an approximate model for faster computation when you have a lot of data or your responses are non-Gaussian.

Implementing Custom Models

The configurability of the above models is limited (for instance, it is not straightforward to use a different kernel). Doing so is an intentional design decision -- we believe that having a few simple and easy-to-understand models for basic use cases is more valuable than having a highly complex and configurable model class whose implementation is difficult to understand.

Instead, we advocate that users implement their own models to cover more specialized use cases. The light-weight nature of BoTorch's Model API makes this easy to do. See the Using a custom BoTorch model in Ax tutorial for an example.

The BoTorch Model interface is light-weight and easy to extend. The only requirement for using BoTorch's Monte-Carlo based acquisition functions is that the model has a posterior method. It takes in a Tensor X of design points, and returns a Posterior object describing the (joint) probability distribution of the model output(s) over the design points in X. The Posterior object must implement an rsample() method for sampling from the posterior of the model. If you wish to use gradient-based optimization algorithms, the model should allow back-propagating gradients through the samples to the model input.

If you happen to implement a model that would be useful for other researchers as well (and involves more than just swapping out the Matérn kernel for an RBF kernel), please consider contributing this model to BoTorch.