Classifier systems: a useful approach to machine learning?
B.G. de Boer · 1994
Classifier systems are sub-symbolic or dynamic approaches to machine learning. These systems have been studied rather extensively. In this thesis some theoretical results about the long-term behaviour and the computational abilities of classifier systems are derived. Then some experiments are undertaken. The first experiment entails the implementation of a simple logic function, a multiplexer in a simple classifier system. It is shown that this task can be learned very well. The second task that is taught to the system is a mushroom-classification problem that has been researched with other learning systems. It is shown that this task can be learned. The last problem is the parity problem. First it is shown that this problem does not scale linearly with its number of bits in a straightforward classifier system. An attempt is made to solve it with a multilayer classifier-system, but this is found to be almost impossible. Explanations are given of why this should be the case. Then some ...