Graphical Causal Model Based Bayesian Optimization for Industrial Process Parameterization

Optimizing production processes is crucial for reducing costs, ensuring product quality, and minimizing downtime. Optimization approaches based on manual parameterization are increasingly reaching their limits, as they depend heavily on the availability of experienced specialists and their process knowledge. Data-driven alternatives such as Bayesian optimization (BO) offer promising solutions, but are limited by their high computational cost and slow convergence on highdimensional problems. In recent years, a number of optimization methods such as CBO, cCBO, MCBO, and DCBO have been proposed that combine causality and Bayesian optimization to solve these challenges. However, these approaches are computationally intensive, sometimes require specialized functions for calculation, either concentrate on a static set of optimization variables or, in the case of dynamic approaches, are heavily dependent on specific assumptions or additional modeling requirements. The approach presented in this work combines Graphical Causal Models (GCM) with BO for parameter optimization. For the efficient and robust calculation of exploration sets, also a path analysis algorithm is developed that analyzes the direct and indirect effects of variables on a target in order to determine relevant variables for the optimization. A real-life industrial case study shows that the approach produces very good and comprehensible results and illustrates its applicability in an industrial environment. By clearly modeling causal relationships and focusing specifically on the most effective variables, the approach offers a resource-efficient and scalable solution for optimizing process parameters with complex interdependence.