A Brutal Intelligence: AI, Chess and the Human Mind | Hacker News Reader

A Brutal Intelligence: AI, Chess and the Human Mind

lareviewofbooks.org

33 pointsrazorburn20 comments

It was the Type B approach — the intelligence strategy — that ended up being the dead end. Despite their early optimism, AI researchers utterly failed in getting computers to think as people do. Deep Blue beat Kasparov not by matching his insight and intuition but by overwhelming him with blind calculation. Thanks to years of exponential gains in processing speed, combined with steady improvements in the efficiency of search algorithms, the computer was able to comb through enough possible moves in a short enough time to outduel the champion. Brute force triumphed. “It turned out that making a great chess-playing computer was not the same as making a thinking machine on par with the human mind,” Kasparov reflects. “Deep Blue was intelligent the way your programmable alarm clock is intelligent.”

Chess engines didn't stop evolving after 1997. Later chess engines are stronger than Deep Blue when running on a laptop or even a smartphone. That's even though they evaluate far fewer positions per second than Deep Blue did. In fact, as of 2014 contemporary chess software running on a smartphone was stronger than chess software from 2006 running on a desktop quad core i7.

http://en.chessbase.com/post/komodo-8-deep-blue-revisited-pa...

http://en.chessbase.com/post/komodo-8-the-smartphone-vs-desk...

Blind speed didn't win the race in the long run. But by the time that was clear, human performance trailed machines by so far that people craving drama from man-vs-machine had lost interest.

I guess non-AI people just cannot understand or acknowledge that thinking is not inherently different than computing.

on which premise do you base this assertion...

it's not inherently different from computing but i don't think any of our current architectures are amenable to "real thinking". a computer that really thinks needs something much, much different from a von neumann or harvard architecture. so much different that it would almost be weird to still call it a computer.

I think brute force (coupled with powerful heuristic of course) is the optimal way for increasing strength of chess play in engines. It has long ago surpassed even the best humans anyway.

What is interesting to me - and really underexplored at the moment - are chess engines that are more human-like, but not for their strength of play. They might be a tad weaker, but much more useful for humans.

Such an engine would be able to annotate (human) games reconstructing most likely intentions of the players - which is what a good human annotator does.

It would therefore be able to teach humans by identifying their specific weaknesses on conceptual level; not merely highlight bad moves.

It could truly play like a human for entertainment or training - current chess engines suck at it, you can weaken them, but their mistakes won't have the same feel as blunders made by human opponents.

Such intelligent chess engines might be able to perfectly simulate playing style of human masters, living or deceased (based on recorded games of course) - so one day we might find a fairly probable answer to how exactly would a 1975 Fischer-Karpov match play out : )

Or generate beatiful, breathtaking games on purpose (there's already been some research devoted to evaluating beauty in chess, eg. https://en.chessbase.com/post/a-computer-program-to-identify...).

Kasparov's philosophy on AI in general is quite interesting. He's also a pretty good speaker with a sense of humor:

https://www.ted.com/talks/garry_kasparov_don_t_fear_intellig...

Nice to see him finally at terms after that one 1997 game--he seems to have been touchy about it even as recently as 2014:

"I think we'll never know unless Kasparov says himself, but you probably won't get to talk to him because he doesn't like to talk about the subject...Kasparov spent years suggesting that IBM cheated, and he hasn't really talked about the game for many years - until now."

http://www.npr.org/2014/08/08/338850323/kasparov-vs-deep-blu...

I'd be very interested to see how the techniques Google used in AlphaGo would fare if applied to chess.

Monte Carlo Tree Search isn't quite suitable for chess, because it's really only a substitute for games like Go which don't have suitable evaluation functions. If you have the ability to write a good evaluation function, as in chess, you don't need to do the "simulation" phase of MCTS. As a result, you basically just use the "tree" phase, but in conjunction with adversarial search heuristics.

Yeah. Just on its face, there's a ton more ambiguity in a go position compared to a chess position. In many cases in chess, there's only one good move and everything else just loses too much material. That seems to be rare in go

That's not really rare in Go.

As someone who plays both games I would say it's less common in Go than in chess. But yeah, by no means is it rare.

They've been doing similar things in Chess for a long time.

For the record, the strategy in Alphago is to pretrain an "intuition" by supervised learning -- they did deep learning on expert games. This is mainly useful for early game moves. For late game play they improved the supervised learning strategy with self-play learning methods specifically MCTS.

In Chess they did the first part by basically stealing knowledge from "opening books" and they do the later game parts with AB pruning.

In Poker they recently did this strategy (look for the "deepstack" poker bot from UAlberta) and were quite succesful. The self play algorithm later on is CFRM. The best poker bot (from CMU) uses CFRM for both the early and late game part with different levels of coarse-graining.

Interesting but I believe that there are events where computers + humans beat just the computers. That shows that there must still be something else that the human adds to the situation that brute force hasn't been able to overcome yet.

umm the brute force more or less learned from the humans, that's how it narrowed down the search space to a manageable level. it's not as if cracked the entire search space all on its own as the search space on its own is quite a bigger than something that even the most powerful computers can tackle.

Yeah, I am having a hard time with the article's black and white distinction of "type A" brute-force chess solvers and "type B" humanlike intelligent engines.

To say that Deep Blue beat Kasparov with nothing but brute-force speed is to neglect the rather intelligent decisions it made, ascribing different weights to pieces in potential sacrifices, positions, development of pieces, control of the center, initiative, and all the other values that a human player uses to play chess.

> To say that Deep Blue beat Kasparov with nothing but brute-force speed is to neglect the rather intelligent decisions it made, ascribing different weights to pieces in potential sacrifices, positions, development of pieces, control of the center, initiative, and all the other values that a human player uses to play chess.

But it's just an evaluation function anyway. A very complex one, but still nothing else. Chess engine doesn't have a notion of a plan (which is how a human approaches chess). All it does is mechanically keep on steering the game towards positions that are evaluted as best by the function.

> All it does is mechanically keep on steering the game towards positions that are evaluted as best by the function.

I wouldn't say so entirely--It's possible to predict multiple moves ahead (before the number of possibilities explodes), and a heuristic is a sort of plan too.

What you're describing sounds like a greedy algorithim more than anything else.

While having a plan is important, it's not the only way a human approaches chess. Experienced chess players develop an ability to evaluate positions intuitively. Without being able to explain fully why and without a specific plan in mind, they're able choose the best position out of several options. In a way, it's very similar to a evaluation function without access to its internal logic.

Chess AIs work by basically brute force with some simple heuristics. Humans still have superior pattern recognition and can play without seeing that many moves into the future at all.

But now the AIs are gaining superior pattern recognition as well. AlphaGo beat expert Go players by training a neural network to recognize good Go moves. Just like how a human would play. And then it combined that with the computers natural ability to search millions of moves. Humans can't add anything to this type of AI.

Reply on news.ycombinator.com