Computational Modeling of Evolutionary Learning.
Roberto R. Kampfner · Deep Blue (University of Michigan) · 1981
A conceptual model of evolutionary learning in the brain, namely, M. Conrad's evolutionary selection circuits model, is studied using a constructive, experimental approach involving computational modeling and simulation. In the conceptual model, networks of formal (enzymatic) neurons develop adaptations based on a process similar to evolution by variation and selection. The behavior of such networks can be changed by modifying the firing properties of the neurons they contain. The firing properties of such neurons are controlled by macromolecules (called excitases) which catalyze events leading to impulse formation in response to specific patterns of dendritic inputs. The main results from the simulations are summarized below. (1) A basic version of the learning algorithm has been developed which yields an effective performance. Four main parameters which determine the performance of this algorithm are: (a) the number of similar networks, (b) the maximum number of networks regarded as "MOST-FIT", (c) the rate of addition of excitases, and (d) the rate of deletion of excitases. (2) Simulation experiments on the learning of pattern classification tasks indicated that single neuron networks learn this type of task more efficiently than redundant networks do. The number of similar networks participating in the learning process helped maintain its efficiency for relatively small increases in the size of the patterns classified. Also, a form of recombination of traits speeds up the learning process. (3) Some simulations indicated that learning tasks interpreted as motor control processes are h and led efficiently by the learning algorithm. (4) Experiments on the sequential behavior of networks indicated that cycle structures defined in terms of the number of state cycles, cycle lengths, and the number of states associated with each cycle are related to the average number of excitase types per neuron in the networks. A middle level of excitase types per neuron gives rise to a richer behavior of the networks, and to a decrease in their structural stability. This level is associated with a larger number of possible states in this type of system. The number of states can be regarded as a measure of the complexity of these systems.