Developing an intelligent framework to play finite-information board games

Computer or video games employ artificial intelligence to provide more interesting and challenging experience. Most AI is achieved by pre-scripted routines. However, designing programs that can play real-life games with human-like ability are a growing area of interest, since the better we become at designing such programs, the more we can learn to apply learning tactics in machines.

The meta-game framework developed by Barney Pell eliminates human bias in learning by generalizing the input it accepts to the rules of a game with a pre-defined class, and designing programs to play them.

Barney Pell uses the concept of “Advisors” – first used by Susan Epstein in HOYLE[1], a key component of his metagame framework derived from extensive human analysis of the class of games, as well as well as expert knowledge. Advisors are resource-limited hierarchical procedures which attempt to compute a decision based on correct knowledge, shallow search and inference.

Although the meta-game framework does a very good job of removing bias, it is limited in its capabilities to search efficiently in a reasonable amount of time to truly demonstrate the use of AI in game-play, because:

1. Manually assigned “weights” to Advisors - In his design of the metagame framework, Pell assigned the weights manually to advisors, via deduction or analysis of the type of game advisor. Even though these advisors are developed for a class of games, there is some potential of bias or human influence creeping in, since the weights assigned to each advisor is potentially the importance the observer places on a certain strategy. For example, “capturing a piece” may be rated higher than “move pawn to convert to King”, in the game of chess.
2. Static Advisors – Both the metagame framework and Hoyle employ pre-determined, “static” advisors, developed with careful human analysis. Such advisors will offer meaningful advice only to games already known, or those that have been used to test the model.

If anyone ever manages to find a working copy of Pell's code, please contact me. This is something I'd be very keen on, however, my search for the code hasn't had much success!

  1. HOYLE
  2. A Strategic Metagame Player for General Chess-Like Games
  3. Entertainment Software Association – Sales & Research Data
  4. Game theory
  5. The metagame project
  6. SAL
  7. A comparison of human and computer game playing
  8. Barney Pell papers FTP site
  9. Nici Schraudolph's go networks
  10. Towards an ideal trainer

No comments:

Post a Comment