This woit investigates the application of evolutionary programming, a multi-agent stochastic search technique, to the generation of recurrent perceptrons (nonlinear hR filters) for time-series prediction tasks. The evolutionary programming paradigm is discussed and analogies are made to classical stochastic optimization methods. A hybrid optimization scheme is proposed based on multi-agent and single-agent random optimization techniques. This method is then used to determine both the model order and weight coefficients of linear, nonlinear, and parallel linear-nonlinear nextstep predictors. The AIC is used as the cost function to score each candidate solution.