An Examination of Lamarckian Genetic Algorithms
Cameron Wellock, Brian J. Ross · 2001
In keeping with the spirit of Lamarckian evolution, variations on a simple genetic algorithm are compared, in which each individual is optimized prior to evaluation. Four different optimization techniques in all are tested: random hillclimbing, social (memetic) exchange, and two techniques using artificial neural nets (ANNs). These techniques are tested on a set of three sample problems: an instance of a minimum-spanning tree problem, an instance of a travelling salesman problem, and a problem where ANNs are evolved to generate a random sequence of bits. The results suggest that in general, social exchange provides the best performance, consistently outperforming the non-optimized genetic algorithm; results for other optimization techniques are less compelling.