Artificial Neural Networks

1. Artificial Neural Networks Michael Prestia COT 4810 March 18, 2008

2. The Human Brain • Composed of neurons • Neurons send signals to each other

3. Neurons • Neurons store and transmit information • Neurons send messages to one another through a synapse http://img460.imageshack.us/img460/8744/neuron6ri.gif

4. Artificial Neural Networks • ANNs simulate neurons to create an artificial brain • The symbol for a synapse in an ANN is a connecting line Anatomical Sketch ANN representation

5. Neural Networks • The brain takes inputs (senses) and produces outputs (actions) • ANNs act in the same fashion Output Output Input Input

6. Example • A driving simulator can take in sensors as input • The output will represent which direction to move Outputs (effectors/controls) Forward Left Right Front Left Right Back Inputs (Sensors)

7. Composition of an ANN • Connections between neurons are called weights • Linearly separable problems do not require a hidden layer Output Layer Hidden Layer Input Layer

8. How Does an ANN Work? out1 out2 • Activation Neuron j activation: H1 H2 w11 w22 w21 w12 X1 X2

9. Recurrent Connections • Allow feedback • Represents a type of memory out Wout-H wH-out H w11 w21 X1 X2

10. Training an Neural Network • Targets can be either known or unknown • Different types of training • Target known • Hebbian Learning • Perceptron Learning • Backpropogation • Target unknown • Neuroevolution

11. Hebbian and Perceptron Learning • Hebbian Learning • Works best when output is independent of input • Simple Hebbian Rule: wi(new) = wi(old) + xiy • Perceptron Learning • changes to match target • wi(new) = wi(old) + αtxi

12. Backpropagation • Designed for at least one hidden layer • General idea • Let activation propagate to outputs • Calculate and assign error values • Adjust weights • Sigmoid activation function is common

13. Backpropagation (cont.) • 5 steps • Calculate error at outputs • Ek = (tk – ok) × ok(1-ok) • Adjust weights going into output layer • Wjk += L × Ek × oj • Calculate error at hidden nodes • Ej = ok(1-ok) × • Adjust weights going into hidden layer • Wij += L × Ej × oi • Repeat

14. Disadvantage of Backprop • Easy to get stuck in local optima http://content.answers.com/main/content/wp/en/6/67/Fitness-landscape-cartoon.png

15. Applications of Backprop • Diagnose medical conditions based on past examples • Learn mouse gesture recognition • Learn to control anything by example

16. Neuroevolution • Uses a genetic algorithm to evolve the weights in a neural network • Genome is direct encoding of weights • Weights are optimized for the given task

17. Code Example

18. Disadvantage of Neuroevolution • Competing Conventions Problem 3! = 6 different representations of the same network A C A B C A B C A B C B A B C B A C

19. Other Types of Neuroevolution • Topology and Weight Evolving Artificial Neural Networks (TWEANNS) • NeuroEvolution of Augmenting Topologies (NEAT)

20. Applications of Neuroevolution • Factory optimization • Game playing (Go, Tic-tac-toe) • Visual recognition • Video Games

21. References • Dewdney, A.K. The New Turing Omnibus. New York: Henry Hold and Company, 1993. • Buckland, Mat. AI Techniques for Game Programming. Cincinnati: Premier Press, 2002. • Machine Learning I: Michael Georgiopoulos • AI for Game Programming: Kenneth Stanley • Forza Motorsport 2 information from Official Xbox Magazine • Neuron image from http://img460.imageshack.us/img460/8744/neuron6ri.gif • Local optima image from http://content.answers.com/main/content/wp/en/6/67/Fitness-landscape-cartoon.png • All other images copied with permission from http://www.cs.ucf.edu/~kstanley/cap4932spring08dir/CAP4932_lecture9.ppt and http://www.cs.ucf.edu/~kstanley/cap4932spring08dir/CAP4932_lecture12.ppt • Forza Motorsport 2 video from http://media.xbox360.ign.com/media/743/743956/vids_1.html

22. Homework Questions • What types of problems do not require a hidden layer? • What are two different methods for training neural networks and how do they differ?