Swim velocity profile identification by using a modified differential evolution method associated with RBF Neural Network
Leandro dos Santos Coelho, Luciano Ferreira da Cruz, Roberto Zanetti Freire · 2013
High level sports require a steady intensification of training in order to raise the athletes' performance. With the purpose of support swimmers and coaches new biomechanical analysis are been performed, becoming one of the most studied areas in swimming. By using technology resources, significant results related to performance improvements are being achieved. Specific analysis of the movements, strength, velocity and projection allow identifying relevant points that directly impact athletes' results. In this context, this work uses a Radial Basis Function Neural Network (RBF-NN) with training combining the Gustafson-Kessel clustering method and the proposed Modified Differential Evolution (MDE) in order to perform the swimmer velocity profile identification. The main idea is to obtain the dynamic of the velocity profile and to use it to improve the athletes' swim style. Differential Evolution (DE) is an evolutionary algorithm that uses a rather greedy and less stochastic approach to solve problems when compared to other evolutionary methods [1]. However, to achieve good performance with DE, the tuning of control parameters is essential as its performance is sensitive to the choice of the mutation and crossover settings. On the other hand, the RBF-NN is a powerful approach for nonlinear identification. This paper combines the two strategies described above proposing a modified DE algorithm based on the association of a sinusoidal signal and chaotic sequences generated by logistic map for the mutation factor tuning. By using data collected from breaststroke and crawl swim style of an elite female swimmer, the validity and the accuracy of the RBF-NN model have been tested by simulations. Results reveal that it is feasible to establish a good model to represent data experimental related to swimming field. Identification results show that MDE outperforms both other tested classical DE approaches for training RBF-NNs in terms of solution quality.