ABSTRACT
Self-adaptive probabilistic neural networks have already been proposed in the literature. Typically, the kernel of the probabilistic neural network is an n-dimensional identity matrix multiplied by the scalar spread parameter, σ. This prevents the network from properly fitting the data. Another approach is to use a diagonal spread matrix, allowing each dimension to assume its own spread value. This approach increases the degrees of freedom of the network and thus allows it to fit better to the available data. However, since the optimization procedure is now multivariate instead of univariate, it is harder and computationally more demanding. To address this optimization problem we employ state-of-the-art computational intelligence optimization algorithms, like Particle Swarm Optimization, Differential Evolution and Evolution Strategies.
