Evolving Neural Networks Applied to Predator-Evader Problem
This document has been relocated to http://scholarsmine.mst.edu/comsci_facwork/248
There were 7 downloads as of 28 Jun 2016.
Abstract
The creation of strategies to meet abstract goals is an important behavior exhibited by natural organisms. A situation requiring the development of such strategies is the predator-evader problem. To study this problem, Khepera robots are chosen as the competing agents. Using computer simulations the evolution of the adaptive behavior is studied in a predator-evader interaction. A bilaterally symmetrical multilayer perceptron neural network architecture with evolvable weights is used to model the “brains” of the agents. Evolutionary programming is employed to evolve the predator for developing adaptive strategies to meet its goals. To study the effect of learning on evolution a self-organizing map (SOM) is added to the architecture, it is trained continuously and all the predators can access its weights. The results of these two different approaches are compared.
