Bayesian Quality Diversity Search with Interactive Illumination
Paul Kent, Juergen Branke · Proceedings of the Genetic and Evolutionary Computation Conference · 2023
This paper presents a novel way for interactively identifying a most preferable solution based on quality and behavioural characteristics. Our algorithm combines the principles of Quality-Diversity Search and Bayesian Optimization to create Gaussian Process surrogate models of the behaviour and fitness space. Unlike traditional Quality-Diversity methods which aim to find good solutions with different behavioural characteristics, we propose a three-step interactive approach that allows a decision maker to efficiently identify the most preferred solution(s). In the first stage, it uses an entropy-based acquisition function to generate an illumination model, followed by an interactive phase where the decision maker can specify regions of interest and a target behaviour. These preferences are then utilized by an improvement greedy acquisition function to guide the optimization process and quickly identify a solution close to the user-specified target. In a case study, with a simulated decision maker, we demonstrate that our approach can find better solutions much more quickly than by selecting the most preferred solution from an archive generated with MAP-Elites.