The paper "Evolution of Neural Networks Using Genetic Algorithms" by Wesley Kerr and Suranga Hettarachchi from the University of Wyoming explores the use of genetic algorithms to evolve neural networks for a simplified RoboCup soccer environment, known as the "Keepaway Testbed." The Keepaway Testbed, created by Peter Stone, involves two teams with equal numbers of players, where one team (keepers) aims to maintain possession of the ball, and the other (takers) tries to take it. This environment provides a realistic simulation of real soccer games, introducing noise to mimic real-world conditions. The authors scaled down the RoboCup challenge to a single agent task: locating a goal in a simulated world. They designed a feed-forward neural network with one input layer, two hidden layers, and one output layer. The inputs included the agent's position, the goal's position, and the agent's heading. The outputs determined the agent's actions: moving forward or turning based on the sigmoid function's output. To evolve the neural network weights, the authors used a genetic algorithm. Each individual in the population was represented as a string of real numbers corresponding to the neural network's connections. Fitness was measured by the number of iterations it took for the agent to find the goal, with higher fitness indicating better performance. The genetic algorithm used stochastic universal sampling for selection, an elitist scheme to preserve the best individual, and recombination and mutation for population variation. After 1500 iterations, the best individual found the goal in 6.25 times the minimum number of steps, achieving a fitness of 0.16. The authors noted that while the fitness values were low, the agent demonstrated intelligent behavior, suggesting that the genetic algorithm successfully learned the neural network weights. However, the project was halted due to the discovery of similar research by Peter Stone, and further exploration of layered learning techniques is planned.The paper "Evolution of Neural Networks Using Genetic Algorithms" by Wesley Kerr and Suranga Hettarachchi from the University of Wyoming explores the use of genetic algorithms to evolve neural networks for a simplified RoboCup soccer environment, known as the "Keepaway Testbed." The Keepaway Testbed, created by Peter Stone, involves two teams with equal numbers of players, where one team (keepers) aims to maintain possession of the ball, and the other (takers) tries to take it. This environment provides a realistic simulation of real soccer games, introducing noise to mimic real-world conditions. The authors scaled down the RoboCup challenge to a single agent task: locating a goal in a simulated world. They designed a feed-forward neural network with one input layer, two hidden layers, and one output layer. The inputs included the agent's position, the goal's position, and the agent's heading. The outputs determined the agent's actions: moving forward or turning based on the sigmoid function's output. To evolve the neural network weights, the authors used a genetic algorithm. Each individual in the population was represented as a string of real numbers corresponding to the neural network's connections. Fitness was measured by the number of iterations it took for the agent to find the goal, with higher fitness indicating better performance. The genetic algorithm used stochastic universal sampling for selection, an elitist scheme to preserve the best individual, and recombination and mutation for population variation. After 1500 iterations, the best individual found the goal in 6.25 times the minimum number of steps, achieving a fitness of 0.16. The authors noted that while the fitness values were low, the agent demonstrated intelligent behavior, suggesting that the genetic algorithm successfully learned the neural network weights. However, the project was halted due to the discovery of similar research by Peter Stone, and further exploration of layered learning techniques is planned.