Improving Adaptability in Optimization-Based Decision Support Systems Through Large Language Models

Since the 1950s, optimization-based Decision Support Systems have emerged as one of the most impactful applications of Operations Research, replacing human-based planning across a wide range of industries. These systems have greatly improved decision-making in terms of efficiency, accuracy, and scalability. However, they also come with a significant limitation: a lack of flexibility and adaptability. Typically, the algorithms used in these systems are predefined during the design phase, following a thorough-and often time-consuming-requirements analysis. They are then hard-coded into the system, making them well-suited for specific, well-understood scenarios. However, this rigidity makes it difficult for such systems to adapt to unforeseen changes or evolving operational needs. As a result, expert intervention, re-analysis, or even a complete system redesign may be required-often at considerable time and cost. This paper explores how Large Language Models (LLMs) can be integrated with traditional optimization algorithms to address this key limitation-lack of flexibility-while preserving their core strengths: speed and solution quality. The core idea is to leverage LLMs to interpret natural language instructions, reconfigure algorithmic components, support preference-based decision-making, and explain the rationale behind their choices. Computational results on a 'rich' Vehicle Routing Problem (VRP) setting-a class of VRPs that incorporate multiple real-world constraints and complexities commonly encountered in last-mile distribution-demonstrate the potential of this hybrid approach.