A NEURAL NETWORK MODEL TO SIMULATE A CONDITIONING EXPERIMENT
Stefano Patarnello, Paolo Carnevali · International Journal of Neural Systems · 1989
The most prominent area where the potential of neural networks is under study is that of engineering applications, such as pattern recognition and speech processing. Another interesting direction which is proposed in the present paper is that of neural networks designed to simulate the interaction of an individual with the external (varying) environment. The simple example presented here consists of a kind of conditioning experiment: a synthetic "organism" (provided with some "sensor" and "motor" neurons which are connected via a neural network) is free to move in a world where it has to find the food for its survival, and is naturally rewarded on the basis of the amount of food it is able to find. The whole training scheme results therefore in a sort of population selection, which finally produces skilled individuals (or networks) able to elaborate relatively subtle search strategies. It is remarkable that the learning scheme produces a differentiation of the task which each neuron performs: some of them behave as "clocks" which resonate at given frequencies (regardless of the external world), while some others are responsible for the interaction with the environment and detect the presence of food.