Co-operation of Biology Related Algorithms meta-heuristic in ANN-based classifiers design
Shakhnaz Akhmedova, Eugene Stanislavovich Semenkin · 2014
Meta-heuristic called Co-Operation of Biology Related Algorithms (COBRA), that has earlier demonstrated its usefulness on CEC'2013 real-valued optimization competition benchmark, is applied to ANN-based classifiers design. The basic idea consists in representation of ANN's structure as a binary string and the use of the binary modification of COBRA for the ANN's structure selection. Neural network's weight coefficients represented as a string of real-valued variables are adjusted with the original version of COBRA. Four benchmark classification problems (two bank scoring problems and two medical diagnostic problems) are solved with this approach. Multilayered feed-forward ANNs with maximum 5 hidden layers and maximum 5 neurons on each layer are used. It means that ANN's structure optimal selection requires solving an optimization problem with 100 binary variables. Fitness function calculation for each bit string requires solving an optimization problem with up to 225 real-valued variables. Experiments showed that both variants of COBRA demonstrate high performance and reliability in spite of the complexity of solved optimization problems. ANN-based classifiers developed in this way outperform many alternative methods on mentioned benchmark classification problems. The workability and usefulness of proposed meta-heuristic optimization algorithms are confirmed.