Backgammon is one of the oldest known board games: a two-player, perfect information game of skill and luck. Players race pieces across the board according to the roll of the die, moving strategically to block their opponent's pieces while advancing their own.
Challenges for AI Research
- The large branching factor (number of different position the game pieces can be in after each turn) makes backgammon a challenging problem AI that can't be solved by simply reasoning through all possible moves.
- Instead, backgammon programs need to recognize patterns in the game and respond to them appropriately
- Programs that use Machine Learning do better than programs that rely on a pre-determined or hard-coded strategy that can only be changed by a person editing the program code
- Temporal-difference learning is a method of machine learning that "learns a guess from a guess", constantly predicting the value of each move and then updating it based on what happens and what it predictions at the next move.
- Hans Berliner's BKG 9.8 was developed in the 1970s and was eventually the first backgammon program to beat a world champion. It used hard-coded evaluations.
- Neurogammon used neural networks and learned from self-play, starting out with random position evaluations and improving them over many thousands of games against itself
- Gerald Tesauro's TD Gammon was the first world-class computer backgammon player. It added temporal-difference learning to Neurogammon
- The First Internet Backgammon Server has been allowing players to play backgammon online since 1992
- Backgammon competitions are a regular part of the Computer Olympiad