A fitness differential adaptive parameter controlled evolutionary algorithm with application to the design structure matrix

Abstract

This paper investigates a methodology for adaptation of the mutation factor within an evolutionary algorithm by means of measuring the improvement differential between successive generations. When no improvement is obtained in an evolutionary algorithm and it has not located the global optimum, it is an indication that the algorithm may have become trapped within a local minimum or maximum. Mutation is a tool within the algorithm that is designed to assist in escaping from these local extremes. It is therefore the premise of this paper that if the preset value for mutation probability is proving insufficient to release the algorithm from entrapment in a local minima or maxima, then a temporary increase in this mutation probability may assist in freeing the algorithm and therefore increasing its chances of ultimately converging on a global optimum. In order to determine when to implement the increase in mutation probability our algorithm measures the fitness improvement between successive generations in the algorithm. When no improvement is detected for a number of successive generations the probability is increased. The design structure matrix (DSM), a scheduling tool that has previously been optimized via the application of evolutionary algorithms, has been used as a practical implementation of differential adaptation to investigate its effectiveness in solving real world problems. Solutions provided by Todd, D. (Multiple criteria genetic algorithms in engineering design and operation, PhD thesis, Department of Marine Technology, University of Newcastle, 1997), are used to benchmark the algorithm’s effectiveness.