Project Title: Leveraging AI Language Models for Automated Generation of Optimisation Models and Solution Approaches
Student: Haolin Zhang
Course: BSc Hons Computer Science
	
Abstract:
Optimisation is widely used in logistics, scheduling, manufacturing, and resource
allocation, but constructing a valid optimisation model from a real-world problem
description remains a specialist task. This project investigates whether large language
models (LLMs) can support the automatic generation of optimisation models from
natural-language descriptions by combining natural language processing, structured
model generation, validation, and solver-oriented export. A modular software
framework was designed and implemented in Python to transform textual problem
descriptions into internal optimisation model specifications and then export them into
Pyomo code. The system includes components for text parsing, prompt construction,
LLM integration, schema-based model representation, validation, and command-line
execution. To ensure reproducibility during development and testing, a deterministic
mock LLM was used alongside an extensible interface for future CodeLlama-based
deployment. The prototype currently supports multiple optimisation problem classes,
including knapsack, assignment, and multi-warehouse transportation problems.
Experimental testing shows that the system can successfully parse semi-structured
natural-language inputs, generate consistent model representations, validate key model
components, and export solver-readable Pyomo formulations. For simpler baseline
problems, the framework also supports end-to-end solution generation through built-in
exact solvers for small instances. The results demonstrate that LLM-assisted
optimisation modelling is feasible when embedded within a structured and validated
workflow, although important limitations remain in relation to unrestricted
natural-language understanding, generic solver execution, and robustness across broader
classes of optimisation problems. The project therefore provides an extensible
proof-of-concept for bridging the gap between informal problem descriptions and formal
optimisation models.