On May 1997 IBM’s Deep Blue, powered by brute computational force, defeated reigning world Chess champion, Garry Kasparov.
I have a very simple idea which I believe could take down Deep Blue — develop a collective chess “game” where a player (Deep Blue, a chess grandmaster, you) competes against the collective wisdom of millions of other chess players.
1. Determine the required number of players to match the skill of a single opponent — you may require 10 people against a good player, 10k for a grandmaster, and 1MM to compete with Deep Blue…
2. Farm out each move to the minimum number of competitors to reach a desired skill level. To play this game, people could simply play a move here and a move there. Not unlike the way a peer-to-peer file sharing system divides work up. To solve the cold-start problem, you could farm moves out to Mechanical Turk in the early days until you have a sufficient number of players.
3. Play around with various methods of weighting individual player’s moves — arithmetic mean, weighting based on some sort of reputation from previous moves, etc.
It’s that simple. You could apply this to Go, Risk, Monopoly — just about any game that requires thinking a few moves ahead.
8 Comments
July 13, 2008 at 6:01 am
Looks like someone is trying it:
http://www.crowdchess.com/
My guess is that the crowd will suck. Though there is an easy way to beat/draw a grandmaster… play 2 at the same time.
Sam
July 13, 2008 at 6:04 am
Very interesting. I think step 3 might be easier said than done though. It’s difficult to isolate the impact a particular move had on the game, especially in real-time. So it’d be difficult to figure out who the “good” people are, and weight them accordingly. Yes, you could use a chess engine (like Deep Blue) to see who’s making good moves, but that defeats the purpose of making WikiChess better than any chess engine.
Then again I would have been skeptical if you told me that an encylopedia that anyone could edit would be so accurate, so take what I say with a grain of salt.
But I’m skeptical of the Wiki concept because I think you’d find that the same few people are consistently proposing the best moves. So why do you need thousands? To truly leverage the crowd wisdom, you’d need specialists (just like Wikipedia). But who wants to be an expert on just King-Knight-Pawn endgames. I want to play the whole game
July 13, 2008 at 11:50 pm
Something akin to this was tried in 1999 with Gary Kasparov, the then reigning world champion, vs the “World”, votes cast by thousands of internet kibitzers:
http://en.wikipedia.org/wiki/Kasparov_versus_The_World
Here’s some interesting commentary about the process:
http://michaelnielsen.org/blog/?p=267
Kasparov won but the World, by all accounts, played very well.
July 13, 2008 at 11:57 pm
In the case of a group-game versus a computer program, I would suspect that the mechanics of allowing a large number of people to think and vote and discuss would handicap the voters. Human grandmasters are already regularly losing against chess programs on microcomputers! Give a Deep-Blue-like machine the chance to calculate for 24-hours a move and it’s unlikely that a million voting humans could defeat it.
You’d have better chances teaming up the top 10 human players and having them confer for 10 minutes a move to limit the computer’s brute-force calculation prowess.
July 14, 2008 at 1:30 am
@jolly
My point exactly. Kasparov wasn’t really playing against the world, he was playing against 4-6 sharp IMs and GMs, whose recommendations were parroted by thousands. That’s why I said “you’d find that the same few people are consistently proposing the best moves”. Irina Krush (who dominated the most of the big high school tournaments I played in), and the other World team leaders were the main reason that the World did so well.
Without the IMs and GMs, “the crowd” would’ve stuggled against Kasparov, let alone Deep Blue.
July 15, 2008 at 10:31 pm
I think you would end up with an excellent filtering mechanism, essentially using the discovered weights for each player as their relative rank. You would find out which of the masses were good at chess. However, there’s no real reason to expect the masses to be good at a specialized activity.
Unfortunately, in much the same way the central limit theorem and the law of large numbers explain large group discovery of accurate estimations, I think you’ll end up with consistently “pretty good moves” for the chess game. However, you won’t be optimizing for the best moves, those that seem perhaps even counter-intuitive when evaluated within the context of a single move.
As an analogy, ask a million people to solve the Monte Hall Problem and take the “most popular” answer. You will invariably get the wrong answer. Ask 10 savvy statistically-minded folks and you’ll be right. The specialized knowledge and strategy required to play “good” chess prevents a million users from doing much more than a mediocre job.
On the other hand, if you could figure out how to parse branches of the game tree to them… limit their necessity to understand an entire game…
July 18, 2008 at 12:58 am
Enjoyed the post AND comments which were quite insightful. Thanks, interesting stuff here.
August 4, 2008 at 1:49 am
This was a thought provoking post. I imagined two teams that can each combine an augmentation approach with collaboration (wiki chess plus one or more Deep Blue machines added to annotate/garden the wiki). Man-machine systems competing against each other. Looks like where we already are in a number of markets with supply chains and ERP systems, except that the capability to foster effective collaboration amongst experts, especially different kinds of experts, seems to be lagging the AI/search/automation techniques.
Leave a Reply