A new concept in the design of finite automata
Frederick Stentiford · ePrints Soton (University of Southampton) · 1974
In recent years considerable effort has been brought to bear upon problems arising in the field of Computer Aided Design. Total mechanization of this design process has generally proved difficult and many successful projects have relied upon a certain measure of human interaction. This thesis proposes and discusses an approach to the automatic design problem in which computer searches are guided by an evolutionary process of random change followed by empirical evaluation. In support of the theory evolutionary methods have been applied to two problems. Firstly a practical method for the reduction of very large finite-state machines is proposed and evaluated. Secondly an evolutionary search is used in the automatic design of an economical set of features for the recognition of OCR B printed characters. Detailed results are presented which give a clear indication of the novel performance o the system. Although the performance falls short of that associated with many commercial equipments, the results are shown to compare favourable with those achieved by an independently and intuitively designed set of features. It is suggested that the work can be extended to Optical Character Recognition problems involving many fonts, and also to problems of fingerprint classification.