Use of Shapley value for selecting centres in RBF neural regressors
The problem of centre selection in radial basis function neural networks (RBFNNs) is re-examined and tackled through a cooperative game theoretic perspective. By resorting to the notion of Shapley value, the approach ranks candidate centres (modelled as game players) for the RBFNN's hidden layer based on a sampled estimation of their marginal contribution to the cross-validation training error. Results achieved on benchmark regression problems are reported, whereby it has been shown that the proposed approach improves on the results delivered by the two well-known algorithms.
