Applying Advanced Learning Algorithms to ALVINN

Abstract

ALVINN (Autonomous Land Vehicle in a Neural Net) is a Backpropogation trained neural network which is capable of autonomously steering a vehicle in road and highway environments. Although ALVINN is fairly robust, one of the problems with it has been the time it takes to train. As the vehicle is capable of on-line learning, the driver has to drive the car for about two minutes before the network is capable of autonomous operation. One reason for this is the use of Backprop. In this report, we describe the original ALVILL system, and then look at three alternative training methods - Quickprop, Cascade Correlation, and Cascade 2. We then run a series of trials using Quickprop, Cascade Correlation and Cascade 2, and compare them to a BackProp baseline. Finally, a hidden unit analysis is performed to determine what the network is learning.