Accuracy-based fitness allows similar performance to humans in static and dynamic classification environments
Adrian R. Hartley · 1999
Traditionally within classifier systems the ability of a classifier to obtain reward (as measured by its strength) indicates the fitness of the classifier within the rule population. However Wilson (1995) proposed a new approach to fitness in terms of classifiers prediction accuracy. This paper presents experiments with two different classifier systems: Newboole (Bonelli et al. 1990) and XCS (Wilson 1995). Both systems demonstrate qualitative matches to data from perceptual category learning in humans. However, the different methods of fitness evaluation of classifiers alter the knowledge the systems learn and maintain. When fitness is based upon strength (Newboole) the system acquires knowledge to solve the classification problem, but when fitness is based on accuracy (XCS) the system acquires a more complete knowledge of the problem space. Further experiments show that the optimal covering map (Kovacs 1997) of knowledge that emerges in XCS allows the system to compensate rapidly in a dynamic classification environment, which is also more similar to human performance on a similar tasks.