Glossary

Arm

Mapping from parameters (i.e. a parameterization or parameter configuration) to parameter values. An arm provides the configuration to be tested in an Ax trial. Also known as "treatment group" or "parameterization", the name 'arm' comes from the Multi-Armed Bandit optimization problem, in which a player facing a row of “one-armed bandit” slot machines has to choose which machines to play when and in what order. [Arm]

Bandit optimization

Machine learning-driven version of A/B testing that dynamically allocates traffic to arms which are performing well, to determine the best arm among a given set.

Batch trial

Single step in the experiment, contains multiple arms that are deployed and evaluated together. A batch trial is not just a trial with many arms; it is a trial for which it is important that the arms are evaluated simultaneously, e.g. in an A/B test where the evaluation results are subject to nonstationarity. For cases where multiple arms are evaluated separately and independently of each other, use multiple regular trials with a single arm each. [BatchTrial]

Bayesian optimization

Sequential optimization strategy for finding an optimal arm in a continuous search space.

Evaluation function

Function that takes a parameterization and an optional weight as input and outputs a set of metric evaluations (more details). Used in the Loop API.

Experiment

Object that keeps track of the whole optimization process. Contains a search space, optimization config, and other metadata. [Experiment]

Generation strategy

Abstraction that allows to declaratively specify one or multiple models to use in the course of the optimization and automate transition between them (relevant tutorial). [GenerationStrategy]

Generator run

Outcome of a single run of the gen method of a model bridge, contains the generated arms, as well as possibly best arm predictions, other model predictions, fit times etc. [GeneratorRun]

Metric

Interface for fetching data for a specific measurement on an experiment or trial. [Metric]

Model

Algorithm that can be used to generate new points in a search space. [Model]

Model bridge

Adapter for interactions with a model within the Ax ecosystem. [ModelBridge]

Objective

The metric to be optimized, with an optimization direction (maximize/minimize). [Objective]

Optimization config

Contains information necessary to run an optimization, i.e. objective and outcome constraints. [OptimizationConfig]

Outcome constraint

Constraint on metric values, can be an order constraint or a sum constraint; violating arms will be considered infeasible. [OutcomeConstraint]

Parameter

Configurable quantity that can be assigned one of multiple possible values, can be continuous (RangeParameter), discrete (ChoiceParameter) or fixed (FixedParameter). [Parameter]

Parameter constraint

Places restrictions on the relationships between parameters. For example buffer_size1 < buffer_size2 or buffer_size_1 + buffer_size_2 < 1024. [ParameterConstraint]

Relative outcome constraint

Outcome constraint evaluated relative to the status quo instead of directly on the metric value. [OutcomeConstraint]

Runner

Dispatch abstraction that defines how a given trial is to be run (either locally or by dispatching to an external system). [Runner]

Scheduler

Configurable closed-loop optimization manager class, capable of conducting a full experiment by deploying trials, polling their results, and leveraging those results to generate and deploy more trials (relevant tutorial). [Scheduler]

Search space

Continuous, discrete or mixed design space that defines the set of parameters to be tuned in the optimization, and optionally parameter constraints on these parameters. The parameters of the arms to be evaluated in the optimization are drawn from a search space. [SearchSpace]

SEM

Standard error of the metric's mean, 0.0 for noiseless measurements. If no value is provided, defaults to np.nan, in which case Ax infers its value using the measurements collected during experimentation.

Status quo

An arm, usually the currently deployed configuration, which provides a baseline for comparing all other arms. Also known as a control arm. [StatusQuo]

Trial

Single step in the experiment, contains a single arm. In cases where the trial contains multiple arms that are deployed simultaneously, we refer to it as a batch trial. [Trial], [BatchTrial]