Honours Project
"An Evaluation of Fast Multi-layer Perceptron Training Techniques for Games"
I adapted this project into a 5 page conference paper which was published on the 6th of September in the proceedings of Game On 2017.
This can be found with the following LINK. A copy of my paper can be downloaded HERE. If you find the idea of the conference paper interesting and have the time, I highly recommend reading my dissertation (download below) as it discusses the project in far more detail.
Abstract
With the rise of Artificial Intelligence (AI) in games leading to the adoption of many academic techniques, multi-layer perceptron neural networks have avoided this trend and rarely been used in a game scenario. This is normally due to long training and development times using the standard error back propagation training technique.
The purpose of this investigation was to compare alternative training techniques to error back propagation in order to see if they can be used to promote the use of multi-layer perceptrons in games.
The application created to serve this purpose was a 2D top down racing game with three different training techniques to control the AI, including error back propagation (EBP), resilient propagation (RPROP) and Random-Minimum Bit Distance Gram Schmidt (RMGS), in which, each training technique was put through three tests. The first was a training test to find the accuracy and training time of the technique. The second was a time trial to test the technique’s ability to race around the track. The final test put the techniques up against a human participant in order to gauge the participant’s opinions on how well the techniques could drive. This allowed for a full comparison between all of the techniques to be made.
Through these tests, it was shown that alternative training techniques, although not as accurate as error back propagation, reduce the training time drastically. The tests also concluded that in a racing game scenario the alternative techniques could also compete with EBP, with the RMGS training technique being the best in every test except accuracy.
This project has shown that multi-layer perceptrons could easily be utilised in game scenarios using these alternative methods and would not require the lengthy training times of error back propagation.