Improving a multiobjective evolutionary algorithm applied to batch scheduling in pharmaceutical manufacturing

Multiobjective Evolutionary Algorithms (MOEA’s) have been developed for optimization problems involving conflicting objectives. Real-world Batch Sequencing (BS) in pharmaceutical manufacturing presents a multiobjective optimization challenge, compounded by constraints and uncertain demands. This complexity often leads to low convergence of feasible solutions. In this study, we evaluate population initialization strategies, propose an optimized mutation operator, and explore various crossover types to enhance solution quality, measured by metrics including the number of non-dominated feasible solutions (NFS), hypervolume (HV), Inverted Generational Distance Plus (IGD+), Error Rate (E), and Coverage of Two Sets (CS).

I. Introduction

Chemical Batch Scheduling (BS) has been investtigated since the early 1990s, considering constraints and real-world uncertainties such as demands and available capacity. Constrained Multiobjective Optimization Problems (CMOP) techniques have gained popularity for schedulling [1] –[5], often using evolutionary algorithms. Given the low convergence of constrained problems, many have employed problem-specific heuristics or operators to enhance the search for feasible solutions, improving both quality and convergence. To address real-world uncertainty Monte Carlo Simulations (MCS) have also been employed to improve probability of solution quality. However, most of these approaches remain unapplied to BS in pharmaceutical manufacturing. Neglecting to establish a robust production planning system can lead to substantial losses, as for the Enbrel pharmaceutical case, with over 200 million USD in lost revenue [4]. MOEAs are well-suited for Multiobjective Optimization Problems (MOPs), processing solutions in parallel and offering problem domain versatility [6].