BoTier in Brief

BoTier provides a lightweight implementation of a composite objective for multi-objective optimization problems, allowing for flexible, hierarchical preferences over both input parameters and measured outputs. It is built as an open-source extension to the BoTorch library, and enables customizable scalarization of objectives, facilitating sample-efficient experiment planning real-world scenarios.

Overview

The core functionality behind BoTier is its hierarchical scalarization method, which allows users to define and optimize multi-objective problems with prioritized preferences across different objectives. BoTier is fully auto-differentiable, enabling seamless integration with gradient-based optimization techniques, and it supports usage as a composite objective that allows for incorporating both input- and output-based objectives.

Key features of BoTier include:

  • Hierarchical Scalarization: A scalarization score for multi-objective optimization that covers tiered preferences across objectives, enabling more nuanced control over trade-offs.

  • Automatic Differentiation: BoTier is fully auto-differentiable, allowing for seamless integration with gradient-based optimization techniques.

  • Integration with BoTorch: BoTier is built as an extension to the widely-used BOTorch library, providing compatibility with a range of optimization and Bayesian inference tools.

  • Composite Objective: In BoTorch’s Monte-Carlo acquisition function optimization framework, BoTier can be used as a composite objective, allowing for flexible incorporation of both input- and output-based objectives.

Overview

Example-Based Explanation

To better understand how BoTier works, we suggest you go through the example at BoTier Tutorial.