Currently, in the geological studies it is clear that the generation process and the dynamics of development of an earthquake belong to the highly nonlinear and non-stationary phenomena. For this reason, in recent years the authors, experts in the development of mathematical models based on Artificial Neural Networks (ANNs), decided to apply these mathematical models to forecast earthquakes. The aim of this experimental study was to test the capability of advanced ANNs and machine learning to estimate the magnitude of the events recorded daily. Features that describe each event are: origin time (UTC), latitude, longitude, depth, and magnitude. With seismic event means an event between 0.1 and 5.9 magnitude, in the database. We have tested the ANN technology on different data sets: a) USGS data from 1976 to 2002; b) USGS and ISIDe data together from 2005 to 2011; c) ISIDe data from 2005 to 2013. This paper aims at demonstrating as the ANNs are a promising technique for earthquake prediction and as an ANN training on the global data on earthquakes is also much more effective for a local earthquake prediction, than an ANN training on local data. In fact, the results show that the ANNs have very good performances both in functional approximation, than in pattern recognition when the training set represents a sample of worldwide earthquakes: 10% of absolute error of magnitude estimation and about 90% of correct classification (1 of 3 classes) in pattern recognition task. The results using only the Italian ISIDe data set are also promising, although the few information available, but less precise than the previous ones: about 99% of correct predictions for events with M <= 2.0, around 75% for moderate events (2.0<M<3.0), and a rate of correct classification between 30% and 40% with events where M >= 3.0. This last result is not surprising, due to the small number of events with this magnitude available in the Italian data set (ISIDe). These results can also be the starting point for the development of a system based on ANNs to provide the daily estimation of possible future seismic events.
