the Bees Algorithm: a novel optimisation tool

This work introduces the Bees Algorithm, a new optimisation algorithm inspired by the foraging behaviour of honey-bees. In its basic version, the Bees Algorithm performs a kind of neighbourhood search combined with global random search and can be used for both continuous and discrete optimisation problems. An improved version of the Bees Algorithm including replacing global random search with interpolation and extrapolation, shrinking neighbourhood size, and abandoning sites with no new information was developed. The improved version could solve benchmark function optimisation problems with less sampling of the search space. The Bees Algorithm has been applied to mechanical design optimisation. Two standard mechanical design problems, the design of a welded beam structure and the design of coil springs, were used to benchmark the Bees Algorithm against other optimisation techniques. Computer-aided preliminary design can be regarded as a special case of optimisation, where the goal is to generate as many solutions as possible above a predefined performance threshold. The higher the number of solutions satisfying the preliminary selection criteria, the greater is the chance to produce a good final solution. An adapted version of the Bees Algorithm for discrete function optimisation was developed and tested on a simple machine design task, preliminary gearbox design. The test consists of finding alternative gearbox configurations that approximately produce the required output speeds using one of the available input speeds. Experimental results show that the Bees Algorithm outperforms random search and a genetic optimisation algorithm. A modified version of the Bees Algorithm was used to search for multiple Pareto optimal solutions in a multi-objective optimisation design problem. Compared to two non-dominated genetic algorithms, the Bees Algorithm was able to find more trade-off solutions. Finally, the Bees Algorithm was employed to train Radial Basis Function (RBF) neural networks for two different problems. Despite the high dimensionality of the problems - each bee represented 2345 parameters in the control chart pattern recognition case and 1581 parameters in the wood defect classification case - the algorithm successfully trained very accurate classifiers. Although the accuracies achieved were marginally lower than those obtained with conventional RBF training methods, the total output errors were less than those for conventionally RBF-trained networks with same number of hidden neurons