Figure 1.

A schematic representation of the torus-shaped (i.e., without edges) virtual world in the evolutionary simulations of den Dulk et al. (2003). Predators (red), food patches (green), and agents (yellow) emit different scents. Agents have receptors for the predator and food scents, predators for the agent scents only. Predators and agents have left and right actuators and move according to the laws of classical mechanics. Collisions between predator and agent result in the agent being “eaten,” between agent and food in the disappearance of the food patch. The predator is controlled by a fixed neural network, so that it moves toward the agent. The agent’s network (an example resulting from the den Dulk et al. simulations is shown at the inset) is subject to mutations and crossovers to its genes (i.e., the connection weights). Starting out with random connection strengths, eventually after hundreds of generations organized behavior (i.e., avoidance of predators, approach of food) emerges, due to the selection of the fittest agents for reproduction.