Artificial Neural Networks

• Artificial Neural Networks (ANN) are frameworks used for elaborating information that take inspiration from nervous biological systems. An ANN is composed by simple units, highly connected with the others: each unit is called neuron, it could receive input signals from the external world or from other neurons and produces an output that could be again either internal or external to the overall network. Each connection node is called synapse and its aim is to modulate the strength of exchanged signals, generally by using a tuneable weight. From this short description, the main characteristics of  ANNs can be noticed:
• parallel computation, i.e. the ability of processing information in parallel thus increasing the speed of elaboration;
• decentralised activity in different units, thus increasing efficiency, reliability and redundancy, but also complexity;
• learning capability is guaranteed by the presence of synapses that allow modifying on-line the strength of connections.

A lot of classes of ANN have been proposed by researchers, and nowadays ANNs are well known both in the academic and industrial domains. Our main objective is to exploit the advantages associated to the utilization of ANNs mainly in control problems: the most relevant feature is adaptivity that could turn out to be very useful when dealing with systems that could undergo any sort of modification. This is the case of the real world and of the space environment, where unexpected events, uncertainties, unknown behaviours, aging phenomena and modifications are frequently present. Moreover, ANNs are characterised by a high level of robustness versus noise and uncertainties. The higher the level of noise is, the lower the performances become, but this feature outperforms classical algorithms that works correctly only under nominal conditions. Parallel calculation guarantees robustness thus allowing fault-tolerance, highly desired in space field.

We have a very deep knowledge of all the classes of ANN, from the traditional feed-forward ones, to the dynamical ANN, until to the Spiking Neural Networks that are the most similar models of biological neurons; we have also used all the learning rules associated to the ANNs from the supervised to the unsupervised ones. The active lines of  research are the following ones:

• theoretical study and comparison of different classes of ANNs and learning paradigms;
• application of the ANNs to a wide class of control problems;
• utilization of evolved dynamical ANNs for the control of a legged-rover for planetary exploration;
• theoretical study on the interaction among learning and evolution of ANNs;
• application of ANNs to Decision Making problems.