Implementation

A simulation of the Baldwin effect consists of two major components: one, a model of the selective forces of evolution, and two, an analogous and biologically justifiable model of learning. The two components that have been used in this simulation are, broadly, the genetic algorithm and the artificial neural network learning to classify a set of input vectors. More specifically, we have used a GA with floating point allele values to model evolution, and a Kohonen Self-Organizing Feature Map [6,4] to model learning.

Throughout the simulation we used a class hierarchy developed by Sutton and Santamaria at the University of Oklahoma [10]. This provided a convenient and clear conceptual paradigm in which the functionality of the simulation was separated into logical components.

The simulation was divided into three parts: the Agent (the collection of neural nets and genetic algorithms), the Environment (the input space of n-dimensional vectors) and a simulation manager that takes care of the communication between the two. Figure 1 [10] gives a graphical depiction of how the simulation was structured.

Figure 1: Agent-Environment Model

The algorithm used for the simulation is described below ignoring most of the low level details of the implementation.

1.

Create a population of m Kohonen Self-Organizing Feature Maps with parameters specified by the user.

2.

Access the input space of n-dimensional vectors.

3.

Create a manager responsible for handling communication between agents and the input vectors.

4.

Run the simulation for a specified number of iterations while the nodes learn to classify the input vectors.

5.

Choose agents that took the fewest number of learning iterations (i.e., the fittest) and create a new population according to the operations of selection, mutation, and crossover. These operations are applied to the initial weights of the neural network (as opposed to the weights as optimized through the Kohonen algorithm).

6.

Record statistics on the old population (average fitness, best fitness, representative chromosome schemas, standard deviation in fitness, etc.).

7.

Repeat steps 3-6 for a specified number of generations.

The hope was that, in the later trials of the simulation, the neural networks would require fewer learning iterations to learn the task, resulting in an increased average fitness among the population. This is what occurred.