Download Project Plan - University of St. Thomas

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Gene expression programming wikipedia , lookup

Transcript
Project Plan
Generic Adaptive Neural Network:
Predator/Prey Scenario
SEIS 761
John Morgan
Motivation
I’m really interested in emergent knowledge. I think that some of the complexities in AI can be
reduced by using AI techniques that use emergent knowledge. Genetic Algorithms and Neural
Networks are both types of AI that use emergent knowledge.
Creating a generic system would give me a toolkit that I could use in the future to create a neural
network to solve any suitable problem.
Goal
Build a Neural Network optimized by a genetic algorithm to chase an object through 2D space.
Description
Phase I
Write an algorithm to create random Neural Networks and write code to host the network and
produce results. The algorithm would use any number of inputs and outputs to create the
network. Build a cost function for unsupervised learning.
Phase II
Build a genetic algorithm to optimize the Neural Network. Optimizations would include:
mutation (changes to activation algorithm, new connections and nodes, modifying the weights of
connections and removing connections) as well as crossover.
Phase III
Run the system to find some optimized networks.
Phase V
Write a UI to display the scenarios.
Phase IV
Analyze the genetic algorithm to find the best mutations.
Phase VI
Modify the algorithm that creates the random neural network to build networks suitable for back
propagation.
Phase VII
Modify the network hosting software to use back propagation. Use back propagation to find
some optimized networks.
Phase VIII
Compare the success of the adaptive neural networks with those found using back propagation.
Success would be determined by number of generations compared to number of cycles of back
propagation and the success of the networks run on similar scenarios.
Phase IX
Create an algorithm to trim the dead nodes from evolved networks.
Tools Required
Visual Studio 2008 (c#)
Knowledge Representation Plan
The Neural Networks would be persisted in a string using Lisp like notation:
(Network
(Input Nodes N1, N2, N3)
(Nodes (N2 (Activation Algorithm)
(Connections N1 N3 N3 . . . )
(Value))
(Output Nodes Nx, Nx+1, Nx+2))
The actual networks would be run in objects like:
Network
Input Nodes
Nodes
Output Nodes
Node
Activation Algorithm
Input Nodes
Weights
Value
Demonstration
Display the UI of a successful Neural Network and graphs of the data.