Search papers, labs, and topics across Lattice.
This paper introduces the agentic simheuristic framework, which uses a large language model (LLM) to coordinate exploratory and exploitative agents within a simheuristic approach for stochastic optimization. The framework addresses the limitations of static control logic in traditional simheuristics and the lack of constrained optimization in LLMs. Applied to the team orienteering problem (TOP) under uncertainty, the LLM intelligently selects diverse solutions from an exploratory agent to seed refinement by an exploitative agent, both using Monte Carlo simulation for evaluation.
LLMs can intelligently coordinate simheuristic components, enabling explainable, AI-driven optimization for complex stochastic problems like the team orienteering problem.
Addressing complex stochastic optimization problems often requires hybridization of search and evaluation methods. Simheuristics combine metaheuristics and simulation but typically rely on static control logic. Meanwhile, large language models (LLMs) offer advanced reasoning but lack robust mechanisms for constrained optimization. We propose the agentic simheuristic framework, a novel architecture that leverages an LLM as a high-level coordinator for simheuristic components. Applied to the team orienteering problem (TOP) under uncertainty, the framework employs an LLM to manage an exploratory agent for broad solution search and an exploitative agent for intensive refinement. Both agents integrate Monte Carlo simulation to evaluate solutions under uncertainty. The LLM guides the process by selecting promising and diverse exploratory solutions to seed refinement, enabling intelligent coordination within simheuristics. We present the framework architecture and provide initial empirical results on TOP benchmark instances, illustrating operational feasibility as a proof of concept and highlighting potential for explainable, AI-driven optimization.
