Expert-guided Bayesian Optimisation for Human-in-the-loop Experimental Design of Known Systems (2312.02852v1)

Abstract: Domain experts often possess valuable physical insights that are overlooked in fully automated decision-making processes such as Bayesian optimisation. In this article we apply high-throughput (batch) Bayesian optimisation alongside anthropological decision theory to enable domain experts to influence the selection of optimal experiments. Our methodology exploits the hypothesis that humans are better at making discrete choices than continuous ones and enables experts to influence critical early decisions. At each iteration we solve an augmented multi-objective optimisation problem across a number of alternate solutions, maximising both the sum of their utility function values and the determinant of their covariance matrix, equivalent to their total variability. By taking the solution at the knee point of the Pareto front, we return a set of alternate solutions at each iteration that have both high utility values and are reasonably distinct, from which the expert selects one for evaluation. We demonstrate that even in the case of an uninformed practitioner, our algorithm recovers the regret of standard Bayesian optimisation.

• The paper introduces a human-in-the-loop framework that combines expert discretion with statistical optimization to enhance experimental design.
• It generates diverse candidate solutions using utility values, predicted outcomes, and visual aids to empower expert decision-making.
• Simulated results show that even partially correct expert choices significantly improve convergence over fully automated Bayesian optimization methods.

Bayesian optimization is a statistical technique used for optimizing functions that are expensive to evaluate. It is commonly used in fields where direct function evaluations require time-consuming experiments or simulations, like material science, bioengineering, and machine learning. A challenging aspect of these applications is that often domain experts hold valuable insights that are not fully utilized in traditional automated Bayesian optimization processes.

The paper introduces a method that involves domain experts more closely in the optimization process, proposing a human-in-the-loop framework for experimental design. This approach is premised on the idea that humans are particularly adept at making discrete choices. The methodology allows experts to impact critical decisions in the early stages of the experiment by selecting from a set of potential solutions presented to them.

In practice, the method generates a series of alternative solutions alongside the optimal solution (in terms of utility) that are well-distributed in the decision space. The decision-maker is provided with a variety of information about these solutions, such as utility values, predicted outcome distributions, and visualization aids, allowing them to apply their domain knowledge effectively. The paper suggests that this process enables domain experts to conduct a form of discrete Bayesian reasoning, where they combine their expertise with the quantitative data provided to make informed decisions on which solutions to pursue.

Experimental results in the paper benchmark the proposed method against standard Bayesian optimization. The authors simulate various types of "practitioner behaviors" to estimate the performance impact of expert involvement. They conclude that even a partially correct expert decision can significantly improve the optimization convergence over purely automated methods. Notably, the method seems to recover the performance of traditional Bayesian optimization when the expert makes decisions at random, indicative of its robustness.

The paper not only brings back the human factor into the optimization loop but also provides a systematic way to harness human intuition in concert with statistical techniques. The authors envision future work to extend their methodology and explore its integration with large-LLMs that might assist or even automate the expert's decision-making step.