Learning in biologically inspired neural networks for robot control

Abstract

Cognitive robotics has focused its attention on the design and construction of artificial agents that are able to perform some cognitive task autonomously through the interaction of the agent with its environment. A central issue in these fields is the process of learning. In its attempt to imitate cognition in artificial agents, cognitive robotics has implemented models of cognitive processes proposed in areas such as biology, psychology, and neurosciences. A novel methodology for the control of autonomous artificial agents is the paradigm that has been called neuro-robotics or embedded neural cultures, which aims to embody cultures of biological neurons in artificial agents. The present work is framed in this paradigm. In this chapter, simulations of an autonomous learning process of an artificial agent controlled by artificial action potential neural networks during an obstacle avoidance task were carried out. The implemented neural model was introduced by Izhikevich (2003); this model is capable of reproducing abrupt changes in the membrane potential of biological neurons, known as action potentials. The learning strategy is based on a multimodal association process where the synaptic weights of the networks are modified using a Hebbian rule. Despite the growing interest generated by artificial action potential neural networks, there is little research that implements these models for learning and the control of autonomous agents. The present work aims to fill this gap in the literature and at the same time, serve as a guideline for the design of further experiments for in vitro experiments where neural cultures are used for robot control.