Fitting distribution-like data to exponential sums with genetic algorithms

Conventional derivative-based algorithms of fitting distribution-like data to exponential-sum functions can be easily trapped in some local minima. This paper is concerned with the development of algorithms of fitting distribution-like data to exponential sums with genetic algorithms. Both binary coding scheme and real-valued coding scheme have been investigated in this research. Experimental results have shown that real-valued coding scheme is more appropriate to the problem of fitting distribution-like data to exponential sums. Testing with real engineering data, it has been demonstrated that the fitting algorithm derived in this paper is quite promising. The fitted exponential-sum models using genetic algorithm can very well describe the measured data. However, for the data with wavy trends, pure exponential-sum functions may not be the best candidate models. More generalized exponential-sum models need to be studied.