Generalization vs. specialization: quantitative evaluation criteria for genetics-based learning systems

In the framework of a novel method for online character recognition adopting a genetic based learning mechanism to design a set of prototypes to be used for the classification, this paper presents a new learning strategy. The proposed strategy is iterative, and it combines two well-known selection mechanism, the roulette wheel algorithm and the elitist approach. Such a combination exhibits both the key features of the two selection mechanisms considered, namely the ability of the former to provide good enough solutions to multimodal problems and the efficiency of the latter in searching for the best solution while dealing with unimodal problem. The first feature is exploited to infer from the samples in the training set, the best set of initial prototypes, while the second one is responsible for producing a new set of prototypes from the initial ones. A method for estimating experimentally when the break-even point between generalization and specialization of the prototypes is reached is proposed as a criterion to terminate the entire learning process