Self-adaptive genetically programmed differential evolution

Differential evolution (DE) is a simple and efficient technique for real parameter optimization over continuous spaces. Its success is highly dependent on the choice of correct trial vector generation strategies and control parameters. Choosing appropriate trial vector generation strategies and control parameters for new problems by trial and error method can be computationally costly and inefficient. This paper proposes a hybrid approach, incorporating genetic programming (GP) with DE, where GP generates trial vector generation strategies based on the problem specification and the state of the population using a simple learning method. Trial vector generation strategies are chosen from this pool of strategies generated by GP. The choice of a particular strategy depends on the type of the problem, initialization values and state of evolution. Consequently, the strategies chosen for different run of the same problem are different. However, it allows self-adaptation to be completely problem dependent and as a result for a unknown problem domain the method is expected to perform better than other state-of-the-art self-adaptive evolutionary techniques. In this method, the control parameter F is eliminated and crossover ratio Cr is evolved with the population and population size NP is still fixed empirically. The performance of this method is extensively evaluated using the CEC2005 contest test instances. Experimental results show that, self-adaptive genetically programmed differential evolution (SaGPDE) leads to quick convergence and produce very competitive results.