ðH geocities.com /chen_levkovich/tdlearningproject.html geocities.com/chen_levkovich/tdlearningproject.html delayed x è`ÔJ ÿÿÿÿ ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÈ PU -A OK text/html °×~ -A ÿÿÿÿ b‰.H Mon, 26 Apr 2004 14:03:16 GMT Ú3 Mozilla/4.5 (compatible; HTTrack 3.0x; Windows 98) en, * è`ÔJ -A
|
TD Learning Project |
We are three students in the Academic college of Tel-Aviv-Yaffo. For our final project under the supervision of Dr. Gideon Dror, we decided to make a self learning system for strategic games using Artificial Neural Networks. Our main reference was Tesauro's article about TD (Temporal Difference) learning and TD-Gammon. We looked for an easy strategic game and after a while we found the BonesGames site. They have a lot of nice games (for free), and we decided to base our game on their LNL game. In order to check our td learning system we first used a kind of random walk game, then a tic tac toe game and only then we started dealing with the real game. Random Walk The board of the game are six squares layered in two rows. 123 456 The player starts the game at square # 1. If the player reaches square # 5 he looses, if he comes to square # 6 he wins. We use the raw data as an input for the system. So this is a very simple game and our system learnt it fine. Tic Tac Toe The famous game. (I guess you all know it). As input we used six parameters: 1. The number of rows and columns that have only one token of the player and no token of the opponent. 2. The same as 1 but with two tokens. 3. The same as 1 but with three tokens. 4,5,6 the same as 1,2,3 but for the opponent. Well the results were good (much better than raw data input). LNL - War in space the real thing. You can see for yourself the results |
Chen Levkovich |
Tali Hildeshaim |
Orly Levkovich-Frank |