Behavioral choice in an electronic nervous system.

Abstract

Much of neural network research being done is performed in simulation, with simple, uniform neuron models and connectivities, and using discrete activity and evaluation steps. This research described herein explores behavioral choices made by a continuous heterogeneous network grounded within a physical robot. Choices studied were dynamically made among both innate and learned behaviors. The study was accomplished by designing and implementing an electronic nervous system composed of neuromimes with excitatory, inhibitory, fatigue and threshold characteristics. The network forms the sole control structure for a tracked vehicle. Inputs arrive via primitive contact, light, and motion sensors, and outputs determine the vehicle's motion. The network incorporates boredom, hunger, and fear drives, as well as the capability for simple conditioning of the touch stimuli to pain reflexes. The choice mechanism is dynamic competition within the network, influenced by current stimuli, internal drives, and recent network history. The resulting network, composed of 87 neuromimes and 342 synapses, utilizes dynamic competition at the sensory interface, among the drives, and at the motor interface. It embodies innate behaviors, learned behaviors, and a method of choice effective over both behavioral sets. To obtain a balance between behavioral persistence and opportunism, we found fatigue to be a critical component in all competitive circumstances. Positive feedback was also found necessary for competition among drive networks. We determined that fatigue and inhibition, while causing similar effects with the neural model, must be handled separately for reasons pertaining to both competition and learning. The inclusion of learning revealed a relation between the structure of plastic synapses and the stability of learned activity. Synapses structured to learn from themselves form permanent memories; those which learn only from other sources of excitation are subject to memory decay on recall trials. We designed and implemented a mechanism for actively reinforcing such actively decaying memory. Finally, it was concluded that learning elements of the network must be placed within the innate competition framework. The implementation presented automatically includes learning and learned behaviors within the innate choice mechanism.