Simulation-based optimization of a production system topology -- a neural network-assisted genetic algorithm

TL;DR

The paper tackles the challenge of topology optimization for production systems by integrating a genetic algorithm with similarity-based mutation/recombination and discrete-event simulation for fitness evaluation. To address high computational costs, it introduces a neural-network surrogate that is evaluated against three architectures (feedforward, pairwise regression, Bayesian) and found to deliver substantial reductions in evaluated designs while maintaining optimization performance, with the feedforward network being preferred for efficiency. The authors demonstrate the approach on an industrial poultry-processing case and a scalability loop-benchmark, showing rapid convergence to optimal designs and strong scalability as the design space grows, especially when using the neural-network surrogate. The work advances practical topology optimization by enabling efficient search over large, highly constrained design spaces, with potential applicability to other domains requiring topology-aware optimization and simulation-based evaluation.

Abstract

There is an abundance of prior research on the optimization of production systems, but there is a research gap when it comes to optimizing which components should be included in a design, and how they should be connected. To overcome this gap, a novel approach is presented for topology optimization of production systems using a genetic algorithm (GA). This GA employs similarity-based mutation and recombination for the creation of offspring, and discrete-event simulation for fitness evaluation. To reduce computational cost, an extension to the GA is presented in which a neural network functions as a surrogate model for simulation. Three types of neural networks are compared, and the type most effective as a surrogate model is chosen based on its optimization performance and computational cost. Both the unassisted GA and neural network-assisted GA are applied to an industrial case study and a scalability case study. These show that both approaches are effective at finding the optimal solution in industrial settings, and both scale well as the number of potential solutions increases, with the neural network-assisted GA having the better scalability of the two.

Figures (14)

• Figure 1: Example of a design space $F$ with $N$ feasible designs. The set of component instances $M$ contains the five possible component instances from four different component types (Infeed, MachineA, MachineB, and Outfeed).
• Figure 2: The poultry fillet distribution system from Paape2023c. In red are component instances and connections which can differ per design. Some components and connections are not present in this specific design. For example, in an alternative design 'assign5' might instead be connected to a 'trim3' component.
• Figure 3: A potential 7-machine loop layout, based on Saravanan2015.
• Figure 4: The chromosome representation of design $N$ as seen in Figure \ref{['fig:design_space_example']}.
• Figure 5: Calculation of the Hamming distances between design $N$ and respectively designs 1 and 2 from Figure \ref{['fig:design_space_example']}. Highlighted in green are bits for which a design differs from design $N$. Design 1 has a higher similarity to design $N$, resulting in a higher chance for it to become the mutated offspring of design $N$.