Man Vs. Artificial Intelligence (AI): From Deep Blue to Deep

[Kirby]

I watched the entire clip - it was pretty interesting. I was amused near the end when Kasparov mentioned that things go in seasons and something along the lines of having yet to see if techniques like neural networks, which were abandoned in the past, would be successful. Demis appeared a bit defensive of AlphaGo's success, even suggesting that it may be successful in chess, too.

Of course, I think Demis is probably correct, but his reaction was defensive IMO.

[pookpooi]

Demis Hassabis defensive probably comes from this article https://motherboard.vice.com/en_us/arti ... atthew-lai

[Baywa]

Real interesting clip. Kasparov is quite a captivating character. He mentioned attempts in the 1970s by Botwinnik ( a former chess world champion) to design learning systems for chess. But they quickly got sidetracked by the brute force systems aka Deep Blue. But then I don't know whether these ideas had anything to do with the concept of neural networks.

Kasparov also thinks that the gap between top chess engines and professional chess players is relatively bigger than the gap between AlphaGo and top go player but that (human) mistakes in go are "deadlier" than in chess. In reply to that Hassabis said "I guess we'll have to put that to the test by teaching AlphaGo to play chess", so Demis brought it up.

[moha]

Kasparov also thinks that the gap between top chess engines and professional chess players is relatively bigger than the gap between AlphaGo and top go player

I kind of see where this comes from, AlphaGo is very new etc, but still seems hard to concretize. The gap measured in what? Percentage of chance of winning a game? I don't know the current state of chess engines - are the best humans' chances of getting a rare draw with white is that close to zero? Alphago did went 60-0...

I also recall the time when chess engines experimented with using more knowledge (instead of more search), before the latter proved more effective. Since go and chess are different, the optimal algorithmic approach can be different as well. But actually I think Alphago's success also owes most to tree search - NNs and MC are just necessary to make that possible at all (pruning the tree to a manageable branching factor, and as a substitute for the evaluation function that's so much easier in chess).

[dfan]

Real interesting clip. Kasparov is quite a captivating character. He mentioned attempts in the 1970s by Botwinnik (a former chess world champion) to design learning systems for chess. But they quickly got sidetracked by the brute force systems aka Deep Blue. But then I don't know whether these ideas had anything to do with the concept of neural networks.

It's been a long time since I read Botvinnik's book, but my recollection is that his ideas for computer chess were mostly based on hand-coded algorithms and rules (as was true for AI in general at the time).