Neural Network

Personal Project (Artificial Intelligence)

Prior to realizing that neural networks are rarely adequate solutions to the problems facing modern games, I undertook the construction of a network class. This neural network was applied to three projects, creating different results in each case.

The key focus of this project was to create a reusable AI component. Construction functions allow for easy creation of multi-layered networks of different sizes, automatic linking, and easy means of changing learning rate, toggling learning on/off, and changing learning activation curves. The network also includes functionality for saving/loading bias values to an XML file.

Rock Paper Scissors-
This neural network was developed initially to learn the precedence of the children's game Rock Paper Scissors. Pitted against the user, a random number generator, and other neural networks, and capable of seeing the oponent's hand, results were "acceptable" at best. The network tended to find trends over several games, rather than react to individual games.

Evasion- (Executable and instructions below)
A single opject in 2D space was taught to evade multiple pursuers. This application was not fully developed, but showed promising results. The network did not perform as hoped when faced with large numbers of pursuers, but it is believed that by including additional data (specifically distance to pursuers), the neural network would have performed well.

Battle Blocks (Tetris clone)-
Applied to a Tetris-like game, the neural network performed very badly. Additional information may have increased it's performance, but it is doubtful that the mushy averaging behavior of the network would have had the precision to properly place the falling blocks.

DEVELOPMENT STATUS:

Completed 2008

VITAL STATS:
Language: C# / XNA
Platform: PC
Dev Time: 1.5 Weeks
Designer: Tim Turner

NOTEABLE FEATURES:
*Easy-to-use interface
*Reusable code base

DEMOS AND CONTENT:

Evasion Demo (Demo software):

 (Requires XNA 2.0) 		Instructions:
The program operates automatically. User may take control of evading ball by toggling the space bar. Toggling user control changes the color of the ball from red to yellow. User can steer the ball by pressing arrow keys. Small dots represent zoomed out view, large dots represent zoomed in view.

top