An Evolutionary Race: A Comparison of Genetic Algorithms and Particle Swarm Optimization for Training Neural Networks.
Brian Clow, Tony White · 2004
This paper compares the performance of genetic algorithms (GA) and particle swarm optimization (PSO) when used to train artificial neural networks. The networks are used to control virtual racecars, with the aim of successfully navigating around a track in the shortest possible period of time. Each car is mounted with multiple straight-line distance sensors, which provide the input to the networks. The cars act as autonomous agents for the duration of the training run: they record the distance traveled and rely on this for fitness evaluations. Both evolutionary algorithms are well suited to this unsupervised learning task, and the networks learn to successfully navigate the course in a minimal number of generations. The paper shows that PSO is superior for this application: it trains networks faster and more accurately than GAs do, once properly optimized. 1. Overview The proposed problem is a vehicle control problem in which virtual cars navigate a course in response to simple distance sensors. Neural networks provide an elegant way to solve this problem – they are very tolerant to noisy data, and so are well suited to problems involving sensory input. Traditional programming techniques are difficult to use without assuming what the proper behavior of a car is. Cars with such assumptions would not be adaptable to different terrains or sensors configurations. Through the use of neural networks any course can be learned no matter what the environment. Sensors can be added or removed without affecting the training process – even completely new types of sensors can be added with no change to the learning algorithm. The trained networks are more tolerant to faults and changes in their design (such as a broken sensor) than a traditional program would be. Many experiments compare the performance of algorithms by running them on a suite of small test functions. The choice of the vehicle control problem investigated here was motivated by a desire to compare algorithms on a test function that could conceivably be used in a real world application. A