This study follows a recent investigation on evolutionary training of counter-propagation neural-networks for multi-objective robot navigation in various environments. Here, in contrast to the original study, the training of the counter-propagation networks is done using an improved two-phase algorithm to achieve tuned weights for both classification of inputs and the control function. The proposed improvement concerns the crossover operation among the networks, which requires special attention due to the classification layer. The numerical simulations, which are reported here, suggest that both the current and original algorithms are superior to the classical approach of using a feedforward network. It is also observed that the current version has better convergence properties as compared with the original one.
展开▼
