From Conceptual Model to Empirical Validation: A Refined Decision Framework for Hybrid Simulation–Machine Learning–Optimization Architectures in Complex Systems

Abstract

Optimizing complex systems, like supply chains and manufacturing, is a tough task. The problem is that detailed simulations are too expensive to use for repeated optimization. To solve this, three main approaches have emerged: using heuristics with simulation, optimization with surrogates, and reinforcement learning. Previous work created a model to choose the right approach based on the problem, but it was just a concept and hadn't been tested. Also, it didn't consider the new role of large language models in optimization. This study fills these gaps. We reviewed research from 2019 to 2025 to understand the key features of complex system optimization and how well each approach works. Then, we proposed a new approach, LLM-Guided Surrogate-Assisted Optimization, which uses a large language model to turn problem descriptions into smart initial sampling strategies, add domain knowledge to surrogates, and narrow down the search space using textual constraints. We mathematically defined this new approach, added it to a decision tree, and tested it in three areas: aerodynamic design, water distribution, and PID controller tuning. Our results show that choosing the right approach based on problem attributes works better than picking one randomly, and our new approach is more efficient when textual domain knowledge is available. This means we can optimize complex systems more effectively, which is a big deal for many industries. By using the right tools for the job, we can make better decisions and improve performance. Our study provides a roadmap for doing just that.