Safe Bayesian Optimization using Interior-Point Methods -- Applied to Personalized Insulin Dose Guidance

TL;DR

This paper introduces a safe Bayesian optimization method that ensures constraint satisfaction by operating within the interior of the safe region, demonstrated on personalized insulin dosing for type 1 diabetes.

Contribution

It proposes a novel interior-point Bayesian optimization framework that guarantees safety constraints are satisfied during learning, applicable with any acquisition function.

Findings

Successfully applied to personalized insulin dose guidance.

Guarantees safety constraints with high probability.

Broad applicability to safety-critical systems.

Abstract

This paper considers the problem of Bayesian optimization for systems with safety-critical constraints, where both the objective function and the constraints are unknown, but can be observed by querying the system. In safety-critical applications, querying the system at an infeasible point can have catastrophic consequences. Such systems require a safe learning framework, such that the performance objective can be optimized while satisfying the safety-critical constraints with high probability. In this paper we propose a safe Bayesian optimization framework that ensures that the points queried are always in the interior of the partially revealed safe region, thereby guaranteeing constraint satisfaction with high probability. The proposed interior-point Bayesian optimization framework can be used with any acquisition function, making it broadly applicable. The performance of the proposed method is demonstrated using a personalized insulin dosing application for patients with type 1 diabetes.