GA-based kernel optimization for pattern recognition: theory for EHW application
M. Yasunaga, Taro Nakamura, Ikuo Yoshihara, J.H. Kim · 2002
An extension of the kernel-based pattern recognition method using a genetic algorithm is proposed. The method is suited to evolvable pattern recognition hardware using FPGAs. In the conventional method one common kernel function is used in the superposition to make discrimination functions. In the extended method each region of the kernel function is optimized individually. For the kernel-region optimization we use a genetic algorithm to solve a large combinatorial problem almost impossible to solve using any brute-force search. A chromosome represents the kernel region in an n-dimensional pattern space, and each locus corresponds to one of the candidates (genes) for an edge length of the kernel region. We have applied the extended method to a sonar spectrum recognition problem and obtained a recognition accuracy of 83.9%, which is much higher than the 62.0% obtained using the conventional kernel-based method and is also better than 82.7% obtained using the nearest neighbor method and the 83.0% obtained using a neural network (backpropagation algorithm). We have analyzed the individually optimized kernel regions and shown that the GA process automatically extracts features in the patterns and embeds the features in the kernel regions.