Natural Computation with Connectionist Systems

This paper presents the evolution of connectionist systems that leads into the so called networks of evolutionary processors (NEPs) and it also shows a general approach to add a learning stage in NEPs. These networks have been proven to be universal models that solve NP-problems in linear time. Most usual disadvantage is that a given NEP only can solve a given problem. NEPs with learning stages can be considered as a more general model to solve several problems, and they are a superclass of NEPs. Some theorems are shown in order to state the computational power of NEPs. First of all, artificial neural networks are revisited (including multilayer perceptrons, Jordan-Elman networks and time lagged networks), then transition P systems and NEPs are shown. Finally, a model of learning in NEPs with filtered connections is proposed