A co-operative parallel heuristic for mixed zero-one linear programming: Combining simulated annealing with branch and bound

This paper considers the exact approach of branch and bound (B&B) and the metaheuristic known as simulated annealing (SA) for processing integer programs (IP). We extend an existing SA implementation (GPSIMAN) for pure zero-one integer programs (PZIP) to process a wider class of IP models, namely mixed zero-one integer programs (MZIP). The extensions are based on depth-first B&B searches at different points within the SA framework. We refer to the resultant SA implementation as MIPSA. Furthermore, we have exploited the use of parallel computers by designing a co-operative parallel heuristic whereby concurrent executions of B&B and MIPSA, linked through a parallel computer, exchange information in order to influence their searches. Results reported for a wide range of models taken from a library of MIP benchmarks demonstrate the effectiveness of using a parallel computing approach which combines B&B with SA.