Tri-phase performance profile of evolutionary search on uni- and multi-modal search spaces

The performance profile for simple evolutionary search is explored over a range of uni- and multi-modal search spaces including an instance of the real-world adaptive distributed database management problem (ADDMP). A three phase cycle is apparent as the mutation rate is increased, in which predominant operation shifts from the simple `discovery´ of new optima, through a `tuning´ phase leading to locally optimal performance, then on to a `transition´ phase as higher order mutations become available. For multimodal problems this cycle can be seen to be repeated over different ranges of mutation rates, whilst for unimodal problems the response profile shows only a single cycle. However as the size of neutral fitness regions is increased on the uni-modal problems, response can be seen to be highly dependent on higher mutation rates. In contrast to the multi-modal search spaces, where crossover can be seen to exploit a larger population size, the performance profile of the uni-modal search spaces is seen to be considerably less affected by population size. The performance profile of the instance of the real-world ADDMP is seen to display characteristics from both the uni- and multi-modal search spaces explored, emphasising the need for both crossover and suitable mutation rates for efficient application of evolutionary search to real-world problems