Evolutionary computation and hybrid systems

Evolutionary computing has been used for many years in the form of evolutionary algorithms (EA)---often mistakenly called genetic algorithms. Normally used for optimisation, EAs suffer from one serious drawback---they are slow. The need for large populations and many generations means that considerable time (and computing resources) are required to solve serious problems adequately. Techniques that 'force' evolution to happen faster will be mentioned and then a technique developed at Swinburne will be described that can speed evolution significantly when fitness evaluation is the major time consuming element in the evolutionary process. There will always be, however, a limit to the improvements that can be made within standard EC and alternate approaches have had to be developed for particularly hard problems. These often involve the encapsulation of information learned within a generation so that it can be used by succeeding generations, thus providing the algorithm with more direction than selection pressure alone provides. A method developed at Swinburne that involves the keeping of historic (Akashic) records for later use will also be described. Developing good intra-generational learning techniques is of limited use unless intra-generational learning is passed on to later generations. Once techniques for the transfer of knowledge between generations are used, the choice of intra-generational learning technique becomes important. Particle swarm optimisation (PSO) can be a good intra-generational learning technique and can be readily combined with evolution to give a superior result to that obtained from either technique alone. Some sample results will be given to support these claims