Learning from heterogeneously distributed data sets using artificial neural networks and genetic algorithms

It is a fact that traditional algorithms cannot look at a very large data set and plausibly find a good solution with reasonable requirements of computation (memory, time and communications). In this situation, distributed learning seems to be a promising line of research. It represents a natural manner for scaling up algorithms inasmuch as an increase of the amount of data can be compensated by an increase of the number of distributed locations in which the data is processed. Our contribution in this field is the algorithm Devonet, based on neural networks and genetic algorithms. It achieves fairly good performance but several limitations were reported in connection with its degradation in accuracy when working with heterogeneous data, i.e. the distribution of data is different among the locations. In this paper, we take into account this heterogeneity in order to propose several improvements of the algorithm, based on distributing the computation of the genetic algorithm. Results show a significative improvement of the performance of Devonet in terms of accuracy.