I have a theory on AI that I would like to write a "whitepaper" about. The distinction I want to explore in AI is learning vs. strategizing. My question is, where can I read other material about this subject?

Let me give a chess example. Let's look at a chess AI as a max-tree, where capturing an enemy unit adds that unit's value to the "move score" for that decision (and likewise losing a piece subtracts that value to the score). Capturing a pawn might net 1 point, a knight 4 points, a rook 5 points, etc.

Strategizing would be AI to apply these points and determine the next move; eg. given ten possible moves, pick the best (max score) at the end of three moves.

Learning would be applying statistical observation to determine those values. If you play 100 games, the AI might decide that capturing a pawn is 2 points, and a knight is worth 7 points, while a rook is only worth 3 points (based on 100 gameplays).

Does this distinction already exist in literature, and if so, where can I read about it?

Edit: Does anyone know a Chess game (with source-code preferably) that utilizes this approach? Maybe Chess960@Home?

  • 3
    \$\begingroup\$ Sounds like a mix of game theory, with 'points' adjusted against epoch based learning. \$\endgroup\$ – deceleratedcaviar Mar 31 '11 at 6:19

What you call strategizing is usually called search in the AI community. It encompasses simple algorithms like A* and DFS, and methods for heuristic design for informed searches like A*.

What you call learning is called machine learning, traditionally split into supervised learning, unsupervised learning, and reinforcement learning. Probably the most important areas to games are genetic programming, neural networks and support vector machines, and Bayesian networks. But machine learning is an enormous field and this is only a small set of the tools it studies.

If you are really interested in the different types of AI approaches, I recommend getting a real textbook, like AI: A Modern Approach rather than reading Wikipedia.

  • 1
    \$\begingroup\$ +1 for A Modern Approach. Great book. Although I disagree with the usefulness per se of neural networks in games (bar Black and White). \$\endgroup\$ – Ray Dey Mar 31 '11 at 15:35
  • \$\begingroup\$ I didn't say they're useful, just important. They've been used in several games and many AI techniques are based on them or compared to them. Unlike, say, data clustering techniques, which I use incredibly often but I don't think I've seen anything more complicated than k-means variations in games. \$\endgroup\$ – user744 Mar 31 '11 at 16:05
  • \$\begingroup\$ That's fair enough, I agree that they are the most applicable areas to games though, they just need a bit of work ;) \$\endgroup\$ – Ray Dey Mar 31 '11 at 16:21
  • \$\begingroup\$ There's a third approach (also "strategizing") called Expert Systems, where you basically find a rule-based algorithm that may require no search trees at all, just essentially a series of if-thens. \$\endgroup\$ – Ian Schreiber Apr 1 '11 at 18:55
  • 1
    \$\begingroup\$ @Ian: I'm familiar with expert systems, but they are not a series of if-thens. In fact modern expert systems are implemented using the tools I described above - one might use machine learning to help gauge possible inference rules, or search using forwards or backwards chaining through those rules. Perhaps you are thinking of decision trees, but even those are often created and tweaked by machine learning and explore multiple paths using search. \$\endgroup\$ – user744 Apr 1 '11 at 21:11

You should definitely read AI a modern approach. The book is a bit expensive but you can't have a serious discussion about AI until you've got some ground work. Also the 2nd edition is as good as the 3rd, so if you're able to find a cheaper 2nd edition take it.

If you really want to get into machine learning, Dr. Mitchell's book has much move indepth information.

It's unfortunate that there is such a large barrier of entry into AI academics. But it wont help you or anyone else if you publish a white paper that uses unique (wrong) vocabulary and discusses techniques already well known in academia.

The field of learning you opponent's behavior to improve your own has several notable entries. Good spam filters do just this. You should look into Paper Rock Scissors AI. What makes PRS unique is that that it's simple and there is no search involved (AKA strategizing). The only way the AI can beat a human is to learn his preferences and exploit them.

Check out this PRS AI bot built by the NYTimes.

  • \$\begingroup\$ Nice, but not what I'm looking for. Joe Wreschnig's answer is essentially what I want -- the terminology of what it is I'm looking to research/write about. Also, I'm not big on terminology and theoretical research; I'd rather write a reusable library and distribute it so people can use it. \$\endgroup\$ – ashes999 Mar 31 '11 at 15:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.