On Evolution of Multi-category Pattern Classifiers Suitable for Embedded Systems
This paper addresses the problem of evolutionary design of classifiers for the recognition of handwritten digit symbols by means of Cartesian Genetic Programming. Two different design scenarios are investigated – the design of multiple-output classifier, and design of multiple binary classifiers. The goal is to evolve classification algorithms that employ substantially smaller amount of operations in contrast with conventional approaches such as Support Vector Machines. Even if the evolved classifiers do not reach the accuracy of the tuned SVM classifier, it will be shown that the accuracy is higher than 93% and the number of required operations is a magnitude lower.
Unable to display preview. Download preview PDF.
- 4.Sekanina, L., Harding, L.S., Banzhaf, W., Kowaliw, T.: Image Processing and CGP. Natural Computing Series, pp. 181–215. Springer (2011)Google Scholar
- 6.Vasicek, Z., Bidlo, M., Sekanina, L., Glette, K.: Evolutionary design of efficient and robust switching image filters. In: Proc. of the 2011 NASA/ESA Conference on Adaptive Hardware and Systems, pp. 192–199. IEEE Computer Society (2011)Google Scholar
- 7.Jin, W., Bin-bin, T., Chang-hao, P., Gai-hui, L.: Statistical method-based evolvable character recognition system. In: IEEE International Symposium on Industrial Electronics, pp. 804–808 (2009)Google Scholar
- 9.Miller, J.F.: Cartesian Genetic Programming. Springer (2011)Google Scholar
- 10.LeCun, Y., Cortes, C., Burges, C.J.C.: The MNIST database, http://yann.lecun.com/exdb/mnist