Google Video: The history of Chess and AI, man vs. machine.

Chess and AI, man vs. machine

Source: Computer History Museum
Video: 2 hr 5 min 57 sec – Jun 22, 2004
URL: http://video.google.com/videoplay?docid=-1583888480148765375

In the early days of Artificial Intelligence, chess was a major focus because it was thought to be a route to understanding problem-solving abilities in general. It was a well defined game in which you could measure the quality of AI play.

A chess champion himself says: “First chess grandmasters came to AI matches to laugh. Secondly they came to watch, but finally they came to learn.”

The prediction of a world-class AI chess player was wrong on timescale. But its a very hard prediction to make. Problem spaces are enormous, ‘searching the entire maze’ is impossible because of cpu power, time and memory constraints. So complete brute force is fruitless. Using pruning (alphabeta) allowed for much more efficient calculations and most algorithms use this technique to reduce the calculation work.
To achieve more, we can wait for processors to become faster, or find a way to use more knowledge. Investing in computer clusters is also an option (they speak of a dichotomy ‘knowledge’ vs. ‘search’). I do believe that each can be improved separately, one doesn’t inhibit the other so the ‘versus’ simply calls for one side to become dominant, not the other to be reduced or left as inferior.
Chess and AI, man vs. machine

Source: Computer History Museum
Video: 2 hr 5 min 57 sec – Jun 22, 2004
URL: http://video.google.com/videoplay?docid=-1583888480148765375

In the early days of Artificial Intelligence, chess was a major focus because it was thought to be a route to understanding problem-solving abilities in general. It was a well defined game in which you could measure the quality of AI play.

A chess champion himself says: “First chess grandmasters came to AI matches to laugh. Secondly they came to watch, but finally they came to learn.”

The prediction of a world-class AI chess player was wrong on timescale. But its a very hard prediction to make. Problem spaces are enormous, ‘searching the entire maze’ is impossible because of cpu power, time and memory constraints. So complete brute force is fruitless. Using pruning (alphabeta) allowed for much more efficient calculations and most algorithms use this technique to reduce the calculation work.
To achieve more, we can wait for processors to become faster, or find a way to use more knowledge. Investing in computer clusters is also an option (they speak of a dichotomy ‘knowledge’ vs. ‘search’). I do believe that each can be improved separately, one doesn’t inhibit the other so the ‘versus’ simply calls for one side to become dominant, not the other to be reduced or left as inferior.

Below are some more (raw) notes of the talk:

Deep Thought
There are various methods used in most chess algorithms:
1. Tuning evaluation function (queen is worth X, rook 5, pawn 1, etc). From material to positional functions. Compare positions (numeric value).
2. Search: what to focus on?
Learning chess the way people do is another acomplisment to be made in the next decades.

Minimax: Calculate moves and countermoves (tree that gets very wide) as long as was computationally feasible.
Alphabeta: Non interesting moves need not be calculated. Reduction of a square root # of moves.
John McCarthy
The amount of branching is moderate. Brute force can win.
The Japanese game ‘Go’ poses a bigger challege for AI, partly because of more branching compared to chess.

Question: How many qbits would it take?
Answer: Still a profound mystery how it can certain solve problems other than faster factoring numbers, searches, etc.

AI separation:
– I don’t know it works well, but it works well.
– Simulation of human cognition.

_Why_ can we beat the hell out of human performance in the game of chess? We still haven’t solved that.

How active is the brain? Unknown.
Mental arithmetic: small part of the brain starts using more energy -> perhaps most of the time most of the brain is inactive. (?! I rather believe that its mostly inactive but can provide important assistance).

Hardware Architectures
VLSI: Very large-scale integration. Massively parallel neural chips in silicon.

Computer performance:
1. Tic-tac-toe
2. Chess (world class performance)
3. Poker (good)
4. Go (moderate performance from decades of man hours of programming)
Time frame answer 1: Upgrading it will require a new idea: when this idea would arise is very hard to predict.
Time frame answer 2: 35 years to Go.
5. ? Crossword puzzles (linguistic challenge)

Interesting questions posed:
Q: What have we learned from AI chess?
A1: That we can solve some problems without having the actual knowledge of the game.
A2: That it’s almost only chess in which AI beats humans.

Cultural difference:

Q: Why not a chess programme that makes mistakes once in a while
A: I put in a losing mode. Over time, the random variation added to the evaluation functions. As the game progresses it starts making bigger mistakes and you start feeling you’re really becomming good.

Issue of generality: #1h37m
Perhaps through the big challenge a lot of research has gone into chess in specific excellence. Perhaps not so much has been learned. General gameplaying agorithm.

We’re looking for a new challenge. Putting a limit on cycles and requiring more intelligence/creativity in a game. You could look for ways that are more analogous to the way humans think.

Language test: Turing test. Line of reasoning: if Einstein were language disabled.

A computer cannot become demoralized?