Obstacle avoidance using a Multi-Layer Perception James Gant & Brett Buehl CS/ECE 539 Fall 2003
Goal Record the actions of a human player and use that data to train a back-propagating neural network to control a vehicle.
Calculating the Neural Network’s Input Find the objects on the Left, Right and in Front of the Vehicle
Calculating the Neural Network’s Input Scale the smallest distance in each region to between 0 and 1. The use that as input for the neural network
Training Data Collection While the user is driving the vehicle, write the distance to the closest object in each region as well as the acceleration/braking/steering information to a file. Use this data to train the neural network to behave like the user did.
Results We were able to train the neural network to slow down and begin to turn when it approached an obstacle but, it would almost always steer to the right which wasn’t always the correct decision. This problem could probably be fixed by using a more complete set of training examples.