A General Modelling System and Meta-Heuristic Based Solver for Combinatorial Optimisation Problems

Abstract

There are many real world assignment, scheduling and planning tasks which can be classified as combinatorial optimisation problems (COPs). These are usually formulated as a mathematical problem of minimising or maximising some cost function subject to a number of constraints. Usually, such problems are NP hard, and thus, whilst it is possible to find exact solutions to specific problems, in general only approximate solutions can be found. There are many algorithms that have been proposed for finding approximate solutions to COPs, ranging from special purpose heuristics to general search meta-heuristics such as simulated annealing and tabu search. General meta-heuristic algorithms like simulated annealing have been applied to a wide range of problems. In most cases, the designer must choose an appropriate data structure and a set of local operators that define a search neighbourhood. The variability in representation techniques, and suitable neighbourhood transition operators, has meant that it is usually necessary to develop new code for each problem. Toolkits like the one developed by Ingber's Adaptive Simulated Annealing (Ingber 1993, 1996) have been applied to assist rapid prototyping of simulated annealing codes, however, these still require the development of new programs for each type of problem. There have been very few attempts to develop a general meta-heuristic solver, with the notable exception being Connolly's General Purpose Simulated Annealing (Connolly 1992). In this research, a general meta-heuristic based system is presented that is suitable for a wide range of COPs. The main goal of this work is to build an environment in which it is possible to specify a range of COPs using an algebraic formulation, and to produce a tailored solver automatically. This removes the need for the development of specific software, allowing very rapid prototyping. Similar techniques have been available for linear programming based solvers for some years in the form of the GAMS (General Algebraic Modelling System) (Brooke, Kendrick, Meeraus and Raman 1997) and AMPL (Fourer, Gay and Kernighan 1993) interfaces. The new system is based on a novel linked list data structure rather than the more conventional vector notation due to the natural mapping between COPS and lists. In addition, the modelling system is found to be very suitable for processing by meta-heuristic search algorithms as it allows the direct application of common local search operators. A general solver is built that is based on the linked list modelling system. This system is capable of using meta-heuristic search engines such as greedy search, tabu search and simulated annealing. A number of implementation issues such as generating initial solutions, choosing and invoking appropriate local search transition operators and producing suitable incremental cost expressions, are considered. As such, the system can been seen as a good test-bench for model prototypers and those who wish to test various meta-heuristic implementations in a standard way. However, it is not meant as a replacement or substitute for efficient special purpose search algorithms. The solver shows good performance on a wide range of problems, frequently reaching the optimal and best-known solutions. Where this is not the case, solutions within a few percent deviation are produced. Performance is dependent on the chosen transition operators and the frequency with which each is applied. To a lesser extent, the performance of this implementation is influenced by runtime parameters of the meta-heuristic search engine.