Reduction of dimensionality for perceptual clustering

Multidimensionality is one of the problems to be solved for a robust methodology in order to be capable of resolving simple and realistic problems. This work establishes a complete methodology based on self-organized maps (SOM) and the expectation-maximization (EM) algorithm that finds an abstract probability function, which is a mix of local experts. An application of this methodology is presented as a case study, where the problem is robot navigation in noisy environments. Readings from seven robot sonars were taken as input for the system, mapped into a two dimension space and grouped into abstract observations, in order to make recognition of navigation space environment dependant and accurate. The goal is to build the capability of predicting observations and of recognizing abstractions that were defined over the data itself