Recommenders

General Information

Recommenders are an essential part of BayBE that effectively explore the search space and provide recommendations for the next experiment or batch of experiments. Available recommenders can be partitioned into the following subclasses.

Pure Recommenders

Pure recommenders simply take on the task to recommend measurements. They each contain the inner logic to do so via different algorithms and approaches. While some pure recommenders are versatile and work across different types of search spaces, other are specifically designed for discrete or continuous spaces. The compatibility is indicated via the corresponding compatibility class variable.

Additional Options for Discrete Search Spaces

For discrete search spaces, BayBE provides additional controls for pure recommenders:

• allow_repeated_recommendations: Controls whether a recommender is allowed to recommend previous recommendations again.

• allow_recommending_already_measured: Controls whether a recommender is allowed to recommend points that have already been measured. This only considers exact matches to the search space.

• allow_recommending_pending_experiments: Controls whether a recommender is allowed to recommend points that have been marked as pending_experiments (see asynchronous workflows). This only considers exact matches to the search space.

Bayesian Recommenders

The Bayesian recommenders in BayBE are built on the foundation of the BayesianRecommender class, offering an array of possibilities with internal surrogate models and support for various acquisition functions.

• The BotorchRecommender is a powerful recommender based on BoTorch’s optimization engine that can be applied to all kinds of search spaces. In continuous spaces, its sequential_continuous flag allows to chose between greedy sequential optimization and batch optimization as the underlying point generation mode. In discrete/hybrid spaces, sequential greedy selection is the only available mode and is thus activated automatically.

  Note that the recommender performs a brute-force search when applied to hybrid search spaces, as it optimizes the continuous part of the space while exhaustively searching choices in the discrete subspace. You can customize this behavior to only sample a certain percentage of the discrete subspace via the sample_percentage attribute and to choose different sampling algorithms via the hybrid_sampler attribute.

  An example on using this recommender in a hybrid space can be found here.

• The NaiveHybridSpaceRecommender can be applied to all search spaces, but is intended to be used in hybrid spaces. This recommender combines individual recommenders for the continuous and the discrete subspaces. It independently optimizes each subspace and consolidates the best results to generate a candidate for the original hybrid space. An example on using this recommender in a hybrid space can be found here.

Clustering Recommenders

BayBE offers a set of recommenders leveraging techniques to facilitate point selection via clustering:

• PAMClusteringRecommender: This recommender utilizes partitioning around medoids.

• KMeansClusteringRecommender: This recommender implements k-means clustering.

• GaussianMixtureClusteringRecommender: This recommender leverages Gaussian Mixture Models for clustering.

Sampling Recommenders

BayBE provides two recommenders that recommend by sampling form the search space:

• RandomRecommender: This recommender offers random recommendations for all types of search spaces. It is extensively used in backtesting examples, providing a valuable comparison. For detailed usage examples, refer to the list here.

• FPSRecommender: This recommender is only applicable for discrete search spaces, and recommends points based on farthest point sampling. A practical application showcasing the usage of this recommender can be found here.

Meta Recommenders

In analogy to meta studies, meta recommenders are wrappers that operate on a sequence of pure recommenders and determine when to switch between them according to different logics. BayBE offers three distinct kinds of meta recommenders.

• The TwoPhaseMetaRecommender employs two distinct recommenders and switches between them at a certain specified point, controlled by the switch_after attribute. This is useful e.g. if you want a different recommender for the initial recommendation when there is no data yet available. This simple example would recommend randomly for the first batch and switch to a Bayesian recommender as soon as measurements have been ingested:

from baybe.recommenders import (
    BotorchRecommender,
    TwoPhaseMetaRecommender,
    RandomRecommender,
)

recommender = TwoPhaseMetaRecommender(
    initial_recommender=RandomRecommender(), recommender=BotorchRecommender()
)

• The SequentialMetaRecommender introduces a simple yet versatile approach by utilizing a predefined list of recommenders. By specifying the desired behavior using the mode attribute, it is possible to flexibly determine the meta recommender’s response when it exhausts the available recommenders. The possible choices are to either raise an error, re-use the last recommender or re-start at the beginning of the sequence.

• Similar to the SequentialMetaRecommender, the StreamingSequentialMetaRecommender enables the utilization of arbitrary iterables to select recommender.

  Warning

  Due to the arbitrary nature of iterables that can be used, de-/serializability cannot be guaranteed. As a consequence, using a StreamingSequentialMetaRecommender results in an error if you attempt to serialize the corresponding object or higher-level objects containing it.