A connectionist neural network model of aggression

Abstract

This thesis tests the biological plausibility of a neural network model of the sensory-motor interface controlling social behaviour in Siamese fighting fish. The model assumes that the sensory-motor interfaces for most motivated behaviours are variations on a standard blueprint. This being assumed, since isolation from releasers of behaviour increases motivation in some cases, it should do so for aggressive display motivation. This was confirmed experimentally in Siamese fighting fish. The experiments also found a transient decremental effect of social isolation. Evidence for the existence of these phenomena in many fish species is discussed. The model provides a mechanistic explanation for such evolutionary conservativism. Another prediction of the model, that presenting sub-threshold social stimuli to isolated animals should increase subsequent display motivation, was confirmed experimentally, as was the surprising, strong prediction that delayed backward conditioning should occur in certain specified circumstances. The model uses connectionism, which is currently the standard mechanism for associative processes, to explain certain reinforcement phenomena. To do this, the usual connectionist synaptic strength change rule was modified by postulating an interference process. The final chapter shows that this alteration actually allows an improved connectionist description of association formation. Association formation and some reinforcement processes, are thus linked together, as due to one process of synaptic strengthening, operating on different network architectures. Finally, the appendices provide a mathematical model for the synaptic strength change rules, a design for autonomous robots with modifiable instincts to perform defined tasks, and a description of how the model explains drive-reduction learning, in a robotics application.