This is a followup to my earlier post describing the use of a simple neural network to control a light following robot. In the original demonstration, the connections between input and output neurons were hard coded with values that were known to steer the robot in the right way. In the current demonstration, the neural network is initialized with random connections and the correct behavior has to be learned.
In the video below, the robot begins with a random 2x2 neural network for controlling the motors based on the values of the two light sensors mounted on the front. A supervised learning algorithm employing the Delta Rule is used to train the network by utilizing a known solution to provide the teaching signals five times per second. At the beginning of the video, you can see that the robot turns away from the light and even goes backward. However, within 10-15 seconds, the network is already sufficiently trained to follow the light beam.
For more information, see http://www.pirobot.org/blog/0006/