Evolving a neural network using dyadic connections

Since machine learning has become a tool to make more efficient design of sophisticated systems, we present in this paper a novel methodology to create powerful neural network controllers for complex systems while minimising the design effort. Using a robot task as a case study, we have shown that using the feedback from the robot itself, the system can learn from experience, or example provided by an expert. We present a system where the processing of the feedback is integrated entirely in the growing of a spiking neural network system. The feedback is extracted from a measurement of a reward interpretation system provided by the designer, which takes into consideration the robot actions without the need for external explicit inputs. Starting with a small basic neural network, new connections are created. The connections are separated into artificial dendrites, which are mainly used for classification issues, and artificial axons, which are responsible for selecting appropriate actions. New neurons are then created using a special connection structure and the current reward interpretation of the robot. We show that dyadic connections can also make an artificial neural network acting and learning faster because they reduce the total number of neurons and connections needed in the resulting neural system. The main contribution of this research is the creation of a novel unsupervised learning system where the designer needs to define only the interface between the robot and the neural network in addition to the feedback system which includes a calculation of a reward value depending on the performance of the robot (or task aim of the system being developed).