Marc Thompson
J9042311

Final Year Project – Machine Learning by Example

I intend to base my final year project on developing an AI system for games. AI systems in games are rarely ever actually intelligent, rather they give the illusion of intelligence by following some pre-set commands without actually displaying any actual signs of intelligence. The reason behind this is the limitations of the hardware of current generation computers and games consoles, current consumer processors simply cannot handle the complex calculations that would be required to make a more intelligent system and perform all the other tasks that are required to run the game.

Rather than programming an AI system that is coded to be able to play a game from the get-go, the system I will create will be coded with the bare minimum of game rules and then will learn how to play and eventually win by playing with a human player. The intention is that this system will be able to be applied to several games with only minor changes to it and still be able to learn to how to win at each game.

I have looked at several research papers written about the field of machine learning, two in-particular should be very useful when creating my client. The first is 'Genetic Algorithms in Search, Optimization, and Machine Learning' by David Goldberg. Although published in 1989 Dr Goldberg's work is still highly relevant and seen as very important in the field of AI. The second paper is 'Data Mining: Practical Machine Learning Tools and Techniques'. This is a more recent publication (January 30th 2011) but it has already been cited numerous times in other AI papers and should give a more up-to-date perspective on the field of machine learning than Dr Goldberg's paper.

I will start by writing two small games using c#, tic-tac-toe and connect four, these games are similar (winning by making a matching row) but with several different rules that separate them for testing purposes. More games may by written if time allows it. For the the AI client, I plan on writing this using the LISP language as it was created specifically for use in AI applications. If there are problems linking the the separate languages then the AI client will probably be rewritten in c# as it will lend itself better to the AI system than LISP will to the game client.

I will be able to test my client in several ways. First I will let a person that has never used the system before, but is familiar with the rules of the game, play against the client that has never played the game before. If the client is learning as it should be, the player will find it more and more difficult to beat the client with each passing game. Another test I can perform is to teach a client how to play the game and then have that client teach another, separate client how to play. By looking at the results of games we should see that second client learns to play, but the first client will also continue to learn from the client it is teaching.