CSCI 5582: Artificial Intelligence

Fall 1996

Semester Project Proposal Excerpts

Projects with a neural net emphasis


Heather Bryant: bryanth@rastro.Colorado.EDU

The project goal is to accurately perform financial forecasting for stocks using a Neural Network. Behavior will be predicted for a minimum of ten stocks, although the end product may include more than ten stocks.


Marcus England: mengland@dimensional.com
Troy Guillory: guillory@cs.colorado.edu

Our aim is to employ an existing neural network software package called "Thinks" by Logical Designs, Inc., on a specific pattern recognition problem in the area of financial forecasting. The goal of the network is to predict a future stock price on the following business day given appropriate training data. A critical feature or attribute of a neural network is in generalizing narrow or specific training data to broader results. Of concern are pitfalls such as overtraining and memorization. We will attempt to find appropriate training data and will contrast and compare the various setup possibilities for the network. This will include research on the algorithms and procedures used, along with their respective results. Architecture and algorithm details are included below.


Jim Teaff: teaff@ucsu.Colorado.EDU

For my final project I shall implement a simple neural-net using MIT Scheme to do character recognition on zip codes similar to the Le Cun et. al. system described in Russell/Norvig page 586. Rather than perform character recognition on handwritten zip codes, I shall implement a system to recognize typewritten zip codes of the form 80013 or 80013-1234 (i.e. the net shall be able to recognize typewritten digits 0 thru 9 as well as a dash "-").


Han Lee: hanlee@tigger.cs.colorado.edu

My project will involve writing a program that does handwritten character recognition using neural networks as described in chapter 19 of Russell and Norvig book (Russell and Norvig 586). Le Cun et al. (1989) have implemented a network designed to read zip codes on hand-addressed envelopes. Instead of recognizing digits, however, my program will try to recognize letters in the alphabet. As in Le Cun et al., the input will consist of a 16 x 16 array of pixels, but the output will consist of 26 output units for upper case letters A through Z. It may contain an arbitrary number of hidden layers and hidden units - the exact number will vary depending on program's performance.


Apiramon Damrongsiri: Apiramon.Damrongsiri@Colorado.EDU
Thammanoon Prathumkaew: FunnyK@aol.com

In this project we will implement the Connect-Four game program by using Scheme language. This program will make a Connect-Four game that is designed to play against humans. One of the problems with the implementation is how to figure out the good evaluation function for Connect-four game. The evaluation function is the function that can tell us good or bad the game-player is doing in any particular situation. We normally use this function with a standard technique to consecutive moves. Unlike most other game of strategy, we propose an alternative approach based on the temporal difference learning, which has been successfully applied to the game of backgammon. With this method, we can train a neural network to be an evaluation function for the Connect-Four game. In this way the game learns to improve itself. This approach comes from the basic idea that "the more the game we play, the better the evaluate function is". Likewise, "the more the game is played, the smarter the computer becomes".


Vlakkies Schreuder: Willem.Schreuder@Colorado.EDU

Use a neural network to interpret Landsat Multi Spectral Scanner (MSS) and Thematic Mapper (TM) satellite images to classify vegetation types of hydrological significance. The ultimate goal is to be able to estimate impacts due to groundwater withdrawals. In order to make these estimates, it is necessary to have a broad classification of the vegetation types. Present techniques of classification available in commercial software does not have sufficient discrimination to distinguish the spectral signature of upland features such as snowfields from valley features such as wetlands.


Back to proposals page