Leveraging human strategy in Bayesian optimization for industrial commissioning

Commissioning industrial processes is often a tedious and difficult task, because certain settings, such as parameters of a controller, need to be adjusted repeatedly until a certain goal is reached. It is also difficult to pass on the information on how to tune a system from experienced operators to junior colleagues, because such knowledge is often tacit and hard to formalize. We approach the tuning problem with Bayesian optimization (BO). We discuss different ways of combining the best of both worlds, i.e., the insights of the expert with the goal-driven exploration of BO. With the goal of using recorded tuning sessions of experts in BO, we dive into two main directions. First, we explore the effect of context changes on a Human First-Computer Last approach, where the expert does the initial rough tuning, and BO does the fine-tuning afterward. Second, we lay out initial steps toward a subgoal-based BO method, based on the observation that humans often break up complex tasks into multiple subgoals to achieve in a given sequence. We evaluate these methods on two simulation setups featuring a cascade controller of a pusher-slider robotic system, and a projectile trajectory planning problem.