File(s) not publicly available
Improving high value manufacturing with simulation-based Bayesian Optimisation
conference contributionposted on 2023-02-23, 04:40 authored by Bruce GunnBruce Gunn, Imali HettiarachchiImali Hettiarachchi, Michael JohnstoneMichael Johnstone, V Le, Douglas CreightonDouglas Creighton, L Preston
This paper presents a novel combination of simulation-based optimisation for the process operation of a physical factory, with an emphasis of understanding how the optimisation results support with design decisions for the evolving system requirements. The study combines a discreteevent simulation (DES) and a Bayesian optimiser. The DES models contain multiple input variables, which can be varied to generate model outputs to help with decision making in stochastic, dynamic environments. To this end, optimising the models output to achieve a maximum or minimum is an integral part of understanding the system performance and guiding design choices. With the absence of a simple (or approximate) mathematical model for describing the system, the simulation models are treated as black-box functions.Sensitivity analysis is a key method for understanding system dynamics of a black-box simulation model. It can be very time consuming and computational expensive, however, when considering complex models with many input variables. Bayesian optimisation (BO) is an attractive alternative which can be used in this context, being sequential and containing self-learning algorithms.In this study we have utilised a model of a manufacturing factory as the optimisation testbed. The model contains many stochastic input variables and operates with day-night shift patterns. A subset of input parameters was optimised in the study, to maximize the factory throughput per day. The optimiser was able to produce good overall throughput results and furthermore, BO results were used to generate charts showing the input-output relationship. This enabled a sensitivity analysis of the factory model, against each of the key parameters used in the optimisation. The results obtained are in line with those observed in practice and helped inform decisions made within the factory. The method deployed here is easily adapted to other models and is easily modifiable for other optimisation techniques.