Approximating McDowell’s Evolutionary Theory of Behavior Dynamics with Stochastic Neural Networks

Behavioral selectionism is the metaphor that learning is like evolution, where successive generations of behaviors develop to increase their demonstrated ability to obtain reinforcement. McDowell’s evolutionary theory of behavior dynamics (ETBD) is a selectionist system based on a sexually reproducing population of bitstrings that undergoes successive rounds of emission, selection, recombination, and mutation. The ETBD is consistent with quantitative behavioral findings under variable schedules of reinforcement. However, it lacks the ability to generalize across high-dimensional input spaces, and it is not biologically plausible. Two neural network implementations of the ETBD are presented, which allow for generalization and hierarchical organization of behaviors. Rather than housing a population of behaviors, these networks encode a population within their synapse weights. Network rules acting on these encoded populations are shown to approximate operations on the ETBD’s explicit populations. The networks are evaluated against twelve quantitative behavioral findings and found to diverge from the results of the ETBD. Genetic drift in the population of behaviors in the ETBD is shown to be responsible for important features of behavior records. Adding a small amount of reinforcement unconditionally at each time step is shown to approximate the effects of genetic drift and leads to convergence between net one and the ETBD’s behavior outputs.