Predation as a Motivation for Schooling Behavior with Simulated Fish

It has been theorized in biology that one of the primary motivations for schooling behavior in fish is escaping predators. When potential prey form larger schools, the predator can become confused and have difficulty trying to target one individual prey from the larger mass. We attempt to model this phenomenon by designing a custom graphical simulator with two different types of simulated robots—prey and predator. The predator is hard-coded to pursue prey that are not in schools and to pursue a random prey at each time step if all prey are in schools. The simulated prey are controlled by a NEAT artificial neural network with fitness based on how many fish are able to avoid being “eaten” by the predator and how long the school as a whole is able to survive. NEAT (NeuroEvolution of Augmenting Topologies) is an evolutionary robotics tool for evolving artificial neural networks (ANNs). NEAT begins with an ANN without any hidden nodes and, as the experiment proceeds, adds hidden nodes as needed. We found that by using a neural network, the prey are able to learn to school successfully within ten generations, even with a small population. These findings suggest that a predator can be good motivation for prey to school and that the schooling task is easier to solve than we originally thought.