• Artificial Intelligence and Classic Games -- Researcher Yves Debilloez has an interesting page covering a variety of the classic games--Connect 4, Checkers, Othello, Chess. It's well worth a look if you're just getting into fiddling with something along those lines....
• Big Blue -- Gerry Tesauro, one of the developers of the infamous Deep Blue chess playing machine, has written an article on the neural network training algorithm his team developed for the machine.
• Silicon Gambits -- An older Discover Magazine article with an emphasis is on computer players of classic strategy games.
• Silicon Champions of the Game -- Another older article, this one in Science News which covers the state of game playing AIs and where they may go in the future.

Twenty Questions:

A game everyone has probably played as a kid, Twenty Questions is particularly well suited to computer play since it is, at heart, a game of narrowing down possibilities.

I suspect there are many implementations of Twenty Questions across the Internet. One of the better ones I've found (thanks to one of this page's fans) is 20Q. 20Q is a great rendition of the game, with the results from each game going into the AI's knowledge base to help make it a better player against the next guy. Very well done...it had a probable guess on the object I had on my desk (a gun) in 16 questions, and it was definite after 30. Very slick.

Mastermind:

• There's a superb approach to solving the classic game of Mastermind using genetic algorithms that's been done by the folks of the GeNeura Team . A variety of papers detailing the solution and algorithms behind the computer player are provided. They also have a version that plays using simulated annealing (a methodology described in the Influence Mapping thread from the comp.ai.games newsgroup in the summer of 1995).



• Stratos has put together another excellent implementation of Mastermind using neural networks to do the heavy lifting. It's a fascinating version that lets you watch the network as it "trains" itself during play. It learns fairly quickly and lends itself to study via the excellent resources on the game's site. Definitely worth a look.

Connect Four, Go-Moku, Qubic (4x4x4 tic-tac-toe):

Solved, by Victor Allis of Vrije Universiteit in Amsterdam, the Netherlands, et al.; all wins for the first player. By "solved" I mean that there exists a program that plays perfectly under tournament conditions (Qubic was proven to be a win for the first player by Patashnik in 1980.)

Some links of relevance for those wishing to learn more:

• For more Connect Four info, click here.
• For more Go-Moku info, try here.
• Finally, you can check out Victor Allis' home page here. Victor has been heavily involved in solving these games as part of his AI research and has written a book on the subject.

Othello (Reversi):

Logistello, written by Michael Buro, is the premier Othello AI on the planet. A good pointer to Logistello information can be found on the site, and the source code was recently made public via the GNU Public License (good job Michael!).

Logistello beat the world champion Othello player, Takeshi Murakami of Japan, in a match held August of 1997. It used a neural network to learn from previous games to improve its knowledge of the game over time and beat Takeshi 6 out of 6 games, a stunning achievement. More info about this epic match can be found at the Logistello site above.

There are a number of excellent pages that focus less on individual Othello programs and more on their inner workings and theory. Two of the better ones:

• www.nada.kth.se/~gunnar/howto.html -- An excellent site that details the most common Othello AI solutions using alpha-beta pruning.
• satirist.org/learn-game/systems/othello/ -- A solid page with many examples of Othello AIs using a variety of AIs and learning algorithms.

Checkers:

• Digenetics is an interesting company I've not heard of until recently, but they seem to be focused on various types of games using neural network technologies. They've got a Checkers game called Checkers With an Attitude that's pretty darn neat.
  
  Looking over the game, it appearss different from many in that they never taught the computer how to play Checkers, they just taught it the basic moves and then let it figure out what worked best. There are at least three distinct personalities you can chose from (represented as rather attractive female AI agents, by the way), and they certainly do seem to have different approaches to the game.
  
  The company also has a Chess game in the works as of this writing, and I expect it will build on the technology here. Kinda fun, and worth checking out so you can see how distinctive the personalities are if nothing else.



• Chinook is likely the best computer Checkers program and was written by a team led by Jonathan Schaeffer of the University of Alberta. Dr. Marion Tinsley, who died 17 April 1995, was either the best or the second best player in the world. Dr. Tinsley beat Chinook 4-2 with 33 draws in 1992; in 1994, Dr. Tinsley withdrew due to illness after playing Chinook to 6 draws; the program then tied Don Lafferty, 1-1, with 18 draws. In 1995, Chinook beat Lafferty, 1-0, with 31 draws.
  
  You can find more information about Chinook at either the University of Alberta's site on the project or at Johathan Schaeffer's site.

Backgammon:

• Snowie:
  
  Using a proprietary neural network technology, Snowie is showing some promise in the computer backgammon race. You can find out more about the software (and order it, if you want) from the company's web site.



• TD-Gammon:
  
  TD-Gammon written by Gerry Tesauro of IBM's T. J. Watson Research Center, is another very strong backgammon program. It is consistently ranked very high in computer-vs-human playoffs and gets better with each iteration.
  
  IBM has a good article on the making of TD-Gammon online here.



• Jellyfish:
  
  Jellyfish is now up to version 3.5. It continues to make use of learning neural networks for its play and even sports a mailing list for its many fans. You can find out more on the official Jellyfish home page.



• HC-Gammon:
  
  Wired magazine recently interviewed Dr. Jordan Pollack of Brandeis University about his backgammon player, based in part of Gerry Tesauro's earlier work. Using both neural networks and genetic algorithms, HC-Gammon is learning to become a better player over time by playing people on the World Wide Web.
  
  You can read the Wired article on Wired's web site, or try your hand at a game here.

Chess:

Deep Blue was regularly ranked as the best computer chess player in the world before going down in history as the first to beat a Master (Garry Kasparov) at the game. You can read an analysis of the 1997 match on the IBM web site or at one of many Kasparov fan sites.

Most commericial chess programs use a variety of interative-searching techniques and manage overall better games than the majority of casual chess players.

Scrabble:

The strongest computer player extant is probably Maven, which debuted in 1987. Maven's tournament results are at least as good as top human players. Additionally, the software is in the hands of dozens of the top human players in North America, and to date nobody has claimed a plus score against it. The program is so strong that the Tournament Scrabble Association requires that Maven "rollouts" be performed on any Maven "tryouts" be performed on any annotated games submitted for publication. Many National Champions believe Maven would have a better chance of winning a Championship tournament than any human, if rules allowed it to compete.

The game has also completely revised the positional theory of the game, having proven to the satisfaction of many players that defense is insignificant. Among the program's contributions is an exact theory of how to evaluate the tiles left in the rack, and an almost perfect endgame player. (When there are no tiles left in the bag, Scrabble becomes a game of perfect information, and Maven uses a search strategy that evaluates such situations better than anyone.)

Hasbro Interactive, Inc., has purchased Maven from its developer. They offer a Windows/Mac CD-ROM version of Scrabble brand crossword game based on it. You can find information about it here.

Go:

In 1995, HandTalk, by Chen ZhiXing, the winner of the Ing tournament at the World Computer Go Congress, won two out of three games against three human experts (all youth champions, 9 or 10 years old)...but HandTalk had a 13-stone handicap. The Ing prize, offered by Ing Chang Chi, for a computer Go program that can beat a Taiwanese professional by the year 2000, is over 40 million Taiwanese dollars (about 1.6 million US dollars).

Go guru Bruce Wilcox has kindly provided a "mega-page" of Go references, rules, AI considerations, etc. This man (and his wife) have analyzed this game to incredible detail, and know it better than just about anybody I think. Take a look and see what you think...

Yet another interesting Go page can be found at the U.S. Go Society's site. This is a bit more of a general topics page, but one can find links to Go-related AI pages there as well.

Recently (as of fall, 2002) I was contacted by some of the folks who put together a pretty nifty game of Go using Java over on their web page. The sourcecode is there as well. Note that the applet does seem to work best under java 1.4, whatever browser you use...

Bridge:

One of the best Bridge programs around is probably Bridge Baron, created by a team led by Tom Throop at Great Game Products. Its bidding and play are at about the level of an intermediate club player, according to its press. You can find more information about Bridge Baron here.

Tignum 2, developed by Stephen J. J. Smith of Hood College, has beaten the Bridge Baron (statistically significantly) at notrump play and may soon be beating it at suit play. I regret that I as yet don't have a good link for more info; I'm working on it.

GIB has claimed to be better than Bridge Baron, having beaten it in several tests conducted by the GIB designers. Check out their web site for more info.

Poker:

Poker is a very hard game to write an AI for, not because it's hard to count cards and play the odds--the best human players can do that (which is why they're kicked out of casinos). No, the biggest and best part of poker is the bluff--that face-to-face interaction between players that just can't happen with a computer game.

There are some noble attempts though, with one of the better ones being the University of Alberta's Poki. Poki is written in Java so you can play over the Internet, and it's kinda neat. The game uses algorithms that cope with probablistic knowledge, guessing (using opponent modeling techniques), and some crude learning based on observation and statistical analysis of each opponent. Pretty neat stuff.

I'm not a big Poker player but I have to say Poki played a good game when I had a go. There's a large support group at the University as well, so you can have lots of interaction with the Poker Research Group there.