This paper analyses the competitive approach\nto the co-evolutionary training of multi-layer perceptron\nclassifiers. Two algorithms were tested: the first opposes a\npopulation of classifiers to a population of training patterns;\nthe second pits a population of classifiers against a population\nof subsets of training patterns. The classifiers are regarded as\npredators that need to ââ?¬Ë?captureââ?¬â?¢ (correctly categorise) the prey\n(training patterns). Success for the predators is measured on\ntheir ability to capture prey. Success for the prey is measured\non their ability to escape predation (be misclassified). The\naim of the procedure is to create an evolutionary tug-of-war\nbetween the best classifiers and the most difficult data samples,\nincreasing the efficiency and accuracy of the learning\nprocess. The two co-evolutionary algorithms were tested on\na number of well-known benchmarks and on several artificial\ndata sets modelling different kinds of common classification\nproblems such as overlapping data categories, noisy training\ninputs, and unbalanced data classes. The performance\nof the co-evolutionary methods was compared with that of\ntwo traditional training techniques: the standard backpropagation\nrule and a conventional evolutionary algorithm. The\nco-evolutionary procedures achieved top accuracy in all classification\nproblems. They particularly excelled on data sets\ncontaining noisy training inputs, where they outperformed\nthe backpropagation rule, and on tasks involving unbalanced\ndata classes, where they outperformed both backpropagation\nand the conventional evolutionary algorithm. Compared\nto the standard evolutionary algorithm, the co-evolutionary procedures were able to obtain similar or superior learning\naccuracies, whilst needing considerably less presentations\nof the training patterns. This economy in the use of training\npatterns translated into significant savings in computational\noverheads and algorithms running time.
Loading....