There is evidence that humans can be more efficient than existing algorithms at searching for good solutions in high-dimensional and non-convex design or control spaces, potentially due to our prior knowledge and learning capability. This work attempts to quantify the search strategy of human beings to enhance a Bayesian optimization (BO) algorithm for an optimal design and control problem. We consider the sequence of human solutions (called a search trajectory) as generated from BO, and propose to recover the algorithmic parameters of BO through maximum likelihood estimation. The method is first verified through simulation studies and then applied to human solutions crowdsourced from a gamified design problem. We learn BO parameters from a player who achieved fast improvement in his/her solutions and show that applying the learned parameters to BO achieves better convergence than using a self-adaptive BO. The proposed method is different from inverse reinforcement learning in that it only requires a good search strategy, rather than near-optimal solutions from humans.
View on arXiv