Deep Blue in Check
By ai-depot | June 30, 2002
Intelligence and Black Boxes
Facts
So, back to Deep Blue. What exactly can it do?
- Given the weight for each feature, Deep Blue can evaluate a board position at an incredibly fast pace.
- Deep Blue can use these evaluations to find the best branch in the search tree, within its horizon.
And that’s pretty much it. Does this make the machine intelligent? Before we get carried away, a quick definition is necessary.
Narrow Intelligence
Until I find the time to complete my long awaited definition of intelligence, this will have to do :P
This is a bit of a controversial issue, but I’m a believer that a intelligence can shine through even while tackling specific problems… or at last most characteristics intelligence can be applied to one non-trivial problem. Flexibility, admittedly, does not enter as a requirement when you only need to play chess! Authors such as Ray Kurzweil call this narrow intelligence.
All I’m argueing is that you can potentially call an entity that plays chess intelligent.
Black Boxes
Some researchers are prejudiced towards intelligence, claiming that it’s only truely ‘intelligence’ if it’s self-aware - like humans. Namely, if you blindly search a game tree for minumum and maximum values, you’re not classed as intelligent (please forgive the random nature of the example ;) Admitedly, minimax shows about as much emergent intelligence as, say, quicksort. It seems a bit strange to call one intelligent and the other not!
Alan Turing, reasercher in computation classed as many as being one of the most influencial figures in A.I., developed a test for intelligence. It took a long time for Turing to develop his test, with a few notable attempts involving chemicals, probes and electrolytes ;) Anyway, it involves putting the machine behind a curtain, and letting it speak to judges. If it can convince them it is human, it passes the test.
This test can also be applied to chess, and instead of placing the candidate behind curtain, picture a black-box instead. In this essay, we’ll conform to this methodology: if an entity in a black box can beat a grand champion, it passes the turing test and we can class it as intelligent (narrow, remember?).
The Team
Deep Blue’s evaluation function is entirely human designed. Tools may have been used to do so (there’s a rumour around that Neural Networks were used in a similar fashion to the Backgammon approach), but essentially it was members of the team that applied their brain to this problem. This was one of the most revealing facts of the guest lecture I was given at York: Murray Campbell spent some of the most fastidiously tedious months of his life tweaking the evaluation function. This process was assisted by chess consultant Joel Benjamin.
This is the key: Deep Blue isn’t intelligent since it alone in the box cannot beat, and could not have beaten the grand champion.
Even if we disregard the training phase, the team had to tweak and modify the machine in between the games. Apart from Kasparov’s erratic openning moves, this explains the slight differences in the behaviour observed during the different games. Without this regular help, an average human grand champion could learn to play against any specific version of Deep Blue, learning over time and finding its weaknesses. The initial training counts for nothing once it has been countered.
So, essentially, if we increase the size of the box, and put Deep Blue and the team inside, we have an intelligent system!
Learning and Adaptation
In order for Deep Blue to class as intelligent in my eyes, it would have needed to learn to update its evaluation function. Quite frankly, I don’t think we’re that far off being able to do this. The technology is available: a combination of neural network learning to learn the weights, along with simple genetic programming to create compound features would allow the machine to adapt to its challenges. The issue of whether it can adapt as quickly as the human brain remains to be seen!
A long shot to getting Deep Blue intelligent would be to learn in a more abstract fashion - rather than based on board features and weights. By analysing game after game of chess, an amazing data-mining algorithm could extract key concepts and express them in such way that the deliberative planning process may reliably draw on them at any stage. But as I said, this will take a lot more effort!
Tags: none
Category: essay |
