Messages

Now we have two options to spend additional computer power. Expanding the tree search and/or expanding the monte carlo simulation. If we expand the tree search to all the leaves of the game tree we basically end up with a min/max approach since the monte Carlo simulation part is not necessary. However, if you use the additional computing power on more monte carlo simulations it does not converge always.

Well, yeah, that's the whole point of MCTS: it provides an approximation of value at the leaves when you can no longer explore the game tree more deeply and outperforms min/max that way. But why would you spend indefinitely more computing power on just the simulation part...? Pointing that out is about as insightful as saying 'imagine a MCTS which only explored one ply deep, clearly it would not do well'; well sure if we assume the most idiotic possible interpretation...

Of course, but not with Monte Carlo tree search. https://en.wikipedia.org/wiki/Monte_Carlo_tree_search It is a heuristic that does not converge to the perfect game.

Yes, it does, if you think about how it works, it converges on perfect play, and your link says so: "it has been proven that the evaluation of moves in MCTS converges to the minimax evaluation" ("with "average" back-up and UCT move selection, the root evaluation converges to the "min-max" evaluation when the number of simulations goes to infinity").

The search tree for a go game is about 300^300 in size. is such a gigantic number, even google wont calculate it.
I think it is about 10^690, compare to the number of atoms in the universe: 10^80

'infinite computing power'; 'in the limit'.

People don't care enough about such matches to make them good PR.

That is why each question will need to be well-funded. So there is a monetary prize for the whistle-blower who reveals most secrets.

Or... you could take that huge heap of money and do things more likely to work, such as feeding it into the market-makers for sports problems and past prediction market standbys like presidential elections (instead of beating an idiosyncratic dead horse).

You aren't capable of the complex calculation of grandmaster play either, so I'm not sure why that would matter. And you can numerically compare the value even more precisely by looking at an AI's estimate for each move.

The estimates the AI give are information we already know. AI is written off of the same strategies I could read in a go book.
There exist a lot of secret strategies in go. People have no incentive to share secrets. These secrets strategies have not been introduced into the AIs yet.

Many such secrets are controlled by fraternities of go players, who pass down the secret for generations.
When a strategy is secret, it is not put through the same scrutiny. So a lot of bad-secrets exist.

Truthcoin would incentivize whistle-blowers to reveal good-secrets, and it would punish people who reveal bad-secrets.
Truthcoin would dramatically increase the amount of accurate open-source knowledge about go strategy.

Huh? It's pretty hard to understand how a Monte Carlo tree 'thinks' since it's taking weighted averages over thousands of leaves of game-trees and seeing their random playouts. You want opacity, you got it. And prediction markets don't incentivize people to 'reveal good-secrets' in your setup - it incentivizes them to reveal good moves. I don't think you've thought this through at all.

For 'easy' PR, we might just steal the popular new items of the day: Where is that Malaysian plane? What will happen with Ukraine? Politics/Sports. etc.

Yes, particularly sports. I have zero interest in professional sports, but I can't deny: there are a *lot* of people who are *very* interested in Bitcoin betting on sports and have been burned by past centralized bookies.

That chess game was played by plurality vote (which works kinda like democracy or majority rule). It is a FAR less effective means of aggregating information from a crowd.
LMSR Prediction markets (like truthcoin), as far as I am aware, are the most effective way of aggregating information from a crowd.

The efficiency of aggregation is irrelevant. People don't care enough about such matches to make them good PR.

Teams of one person with an AI are much better at go than either humans or AI alone. I suspect that such teams will be the biggest winners in this experiment.

Advanced chess is already on the outs: the chess AIs have gotten so good that you need to be pretty much a grandmaster to add much of anything. In another decade I suspect it'll be toast.

I don't think games played with monte carlo trees are very interesting. I don't get any better by watching computers do this, because I am incapable of such complex calculation.
Watching people bet would be incredibly interesting to me because on every turn you get a weighted graph of the entire board. You can numerically compare how good every move on the board is based on the price of the shares.

You aren't capable of the complex calculation of grandmaster play either, so I'm not sure why that would matter. And you can numerically compare the value even more precisely by looking at an AI's estimate for each move.

Actually the capability of Monte Carlo tree search are provable limited. Infinite computing power would not lead in all cases to the perfect game.

Cite? My understanding was that in the limit it searched the full game tree and so played optimally.

If we asked the prediction market which move is best, and each move was well funded, we could play the best game of go that has ever been played in history.

Existing Go AIs are embarassingly-parallel Monte Carlo trees with anytime properties where the more computing power you invest the better each move is, and the ratings are based on relatively quick matches (maybe an hour per move at most?); so if anyone cared about creating the best game ever, all you have to do is is buy something like a 64-core machine and run two copies against each other for a year or three. That no one has done this, or has repeated 'Kasparov vs the Internet', suggests there's not a whole lot of interest in such games. As Hanson might 'gameplaying isn't about winning'.

You are either holding one of two insane beliefs: that the shooting was planned, but didn't leak (which it would), or leaked in a stupid way (mis-labled dates on [unreliable] internet searches), or that the shooting didn't happen at all (which is, for all practical intents and purposes, impossible).

Highly unreliable. When I was researching various Google services and software products for my analysis in http://www.gwern.net/Google%20shutdowns , I made heavy use of the timeline restriction tool, and there were bogus search hits *all the time*. It was frustrating as heck.

People backdate posts, people forward date them, people remix posts... Pages had to be read very carefully, crosschecked against the Internet Archive, etc. You may ask 'why would anyone backdate a post about Sandy to before Sandy?' but there are literally hundreds of millions of people (many quite stupid or careless) out there messing around with stuff, all software has bugs, there's plenty of weird projects running for purposes which are arcane or totally opaque (I recently read about something like 50k videos uploaded to YouTube consisting of moving shapes; a modern-day numbers station? no, ultimately turned out to be just some DVR or consumer electronic testing out its video upload functionality), and over the hundreds of billions of webpages, there's all sorts of weird stuff.

Code Select Expand

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA512

[quote author=toast link=topic=26.msg113#msg113 date=1401939828]
sorry to derail but
holy fucking shit are you THE gwern? Big fan here...
[/quote]

Yes, it's me.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1

iEYEAREKAAYFAlOQkO4ACgkQvpDo5Pfl1oIuqQCfWZ3qEmTRfZzHXo3y+lRLDk1u
/sMAn1/rPQfCTd2TLc0fS2axK1GvjTIu
=NfLt
-----END PGP SIGNATURE-----


It's a waste of time. What's the use-case for multi-sig on a prediction market, exactly? Truthcoin isn't intended to be a general-purpose currency, you're not going to be buying many products where you're worried you'll be ripped off. Why do you need split control of your personal shares in each market? I mean, there's some weak security arguments, but that's not enough to make it any sort of top priority. Multi-sig is the sort of luxury that should be added long after all the core functionality is up and running and debugged and refined for end-users.