Playing against computers, do neural nets change things?

[sparky314]

Advice is always given for people to stop playing against computers and play against other humans, because computers can be taken advantage of, and they play haphazard moves due to the algorithms behind them. I've also given this advice to beginners.

But with the advent of AlphaGo and the neural nets, computers are going to play much closer to human level play. How long before this advice is no longer valid? Especially with Crazy Stone's new version release in just a few days. While people should still play people, will they also gain an advantage by playing a stronger, human-like computer?

Should we welcome our robot overlords?

[Kirby]

I think people have always gotten benefit from playing computers. I went from about 15k to about 5k primarily because of GnuGo.

[Mike Novack]

" because computers can be taken advantage of, and they play haphazard moves"

Look, we shouldn't make generalizations like that. And because what conclusions we can make about the strengths and weaknesses of computer programs depends on the level of go required, I am far from certain that neural nets necessarily change the picture much (maybe they do, but too soon to tell).

a) Go knowledge based AIs can't do much better than perhaps 5 kyu and they are subject to systematic weakness (something you could take advantage of). But they normally would not play haphazard moves.

b) MCTS programs are useful in the range above that up to 2-3 dan on ordinary hardware (6-7 dan on powerful machines. Some of them might make what seem like haphazard moves, others do not, but remember, a 7 kyu playing a 2 dan HUMAN at nine stones might think some of the moves made by the stronger player were "loose" when they were not. But they can't really be weakened much below 3-4 kyu without becoming erratic. They do not have weaknesses easy to exploit.

c) We won't know much about neural nets as opponent until there is more playing experience. I am referring to when using modest hardware. Obviously the best of these playing on really powerful machines are up there with top human players. But how weak can they be made to play an still work? That might require special training (training using game records of players of more modest abilities)

NEWS FLASH --- farther down in the forum Rémi has announced about to release a new version of CrazyStone with the lower levels pure neural net with the net trained for different levels. Does appear to play human like. The haphazard is the NORMAL haphazard of a weaker player.

[Calvin Clark]

For anyone who is not yet high dan, I think the advice to avoid computers is already questionable for people who are willing and capable of putting down an amount money equivalent to the price of a handful of books. (I know there are a number readers who will only use freeware, and my answer there might be to wait a little bit, whether that be for stronger free programs or for free games on KGS against strong bots.)

That being said, most people judge their progress by their games against humans and I think that should still be the gold standard. You have to be able to refute weird stuff that players around your level try and computers are unlikely to provide those kinds of varied tests. Consider some of the bizarre opponents you've had! If you can only play well against "good" players or a particular set of "bad" players, that's less desirable, IMHO.

Anyway, our friends in the chess world have had to live with this situation for some time and may have some useful things to say about the place of computers. I think most of it is along the lines of "if you use them, don't stop thinking for yourself."

[Krama]

For anyone who is not yet high dan, I think the advice to avoid computers is already questionable for people who are willing and capable of putting down an amount money equivalent to the price of a handful of books. (I know there are a number readers who will only use freeware, and my answer there might be to wait a little bit, whether that be for stronger free programs or for free games on KGS against strong bots.)

That being said, most people judge their progress by their games against humans and I think that should still be the gold standard. You have to be able to refute weird stuff that players around your level try and computers are unlikely to provide those kinds of varied tests. Consider some of the bizarre opponents you've had! If you can only play well against "good" players or a particular set of "bad" players, that's less desirable, IMHO.

Anyway, our friends in the chess world have had to live with this situation for some time and may have some useful things to say about the place of computers. I think most of it is along the lines of "if you use them, don't stop thinking for yourself."

The difference between chess engines like Stockfish and Komodo and AlphaGo is huge. Chess computers really play inhuman moves most of the times and you really need to be careful when studying computer games. Chess engines are mostly used to check for blunders and on high levels for smaller mistakes in tactics.

While AlphaGo did play some unusual moves (some were small mistakes) it did play quite humanlike.

I think humans will be able to learn from go programs even more than chess player learn from chess engines.

[dfan]

Chess computers really play inhuman moves most of the times and you really need to be careful when studying computer games. Chess engines are mostly used to check for blunders and on high levels for smaller mistakes in tactics.

I think this is overstating the case. Chess engines play human moves most of the time, though of course they will make the occasional "computer move", and when following an engine suggestion it is very important to make sure that you understand the followups so you don't get caught out by the fact that the validity of a move depends on a precise odd-looking response three moves into the variation tree. Engines are routinely used to check for positional errors as well as tactical ones, though their recommendations have to be studied with care.

[erislover]

I don't think there is anything wrong with playing computers in the abstract sense. A 6k computer makes bad moves and a 6k human opponent makes bad moves and if you never study your play it won't matter whether you are playing GNU Go or not. There are few bad habits you would acquire from a computer that you wouldn't also acquire from a human, if you played both the same.

The problem is that some of the faster bots like Aya or GNU Go can turn a game with no time settings into blitz, whereas a human is likely to take his or her time at certain key points to try to read, if that time is available. You can try taking your hand off the mouse after every turn but it's still difficult psychologically to spend time when your opponent isn't. So bots can encourage unthinking play.

I don't think even weird center-oriented bots like pachi or mogo are bad. If the bot makes bad moves, prove they are bad through play and you will learn something; if you could not prove it, maybe it wasn't bad at all and you also learned something.

I think humans will be able to learn from go programs even more than chess player learn from chess engines.

Well I can't speak about Alphago vaporware but crazystone has a similar vibe to it. It's analysis is very interesting and I love looking at my own games with it, but you have to be able to pull a Jerry from chessnetwork and say, "That's computer go, I wouldn't play this way." On the other hand, you have to be strong enough to make that judgment in the first place, so there's some bootstrapping problem here (Dunning-Kruger).

[shapenaji]

I wonder if it's possible to train a policy network on kyu players in order to get bots that are useful for people learning.

[hyperpape]

I wonder if it's possible to train a policy network on kyu players in order to get bots that are useful for people learning.

I believe Mike is saying that Remi (Crazystone) is working on just that.

[shapenaji]

Fair enough, wow, I somehow managed to read right over that. That's awesome.

I'm really curious then about how strong the computer can become playing against itself and optimizing. It would seem to say something very interesting about people learning by playing against people their same strength.

[Uberdude]

I wonder if it's possible to train a policy network on kyu players

shapenaji, you may like this post with an example game from the kyu-trained Crazystone from RemiC.

I'm really curious then about how strong the computer can become playing against itself and optimizing. It would seem to say something very interesting about people learning by playing against people their same strength.

Doesn't AlphaGo answer that question? Train on KGS 1-9d game data (I'm still not clear if they used GoGoD or go4go with pro games as training data for Fan Hui or Lee Sedol version, there have been conflicting reports), play yourself and improve and then beat Lee Sedol. Deepmind have said they are also thinking about training a new bot from just the rules, to see if self-play without the initial guidance of human experts can reproduce master level skill (this is the approach they used on the Atari computer games playing AI); that would obviously take far longer but would also be interesting to see if it rediscovers all our standard opening theory, josekis etc or plays a unique style (wacky centre moyos?). That approach has deep impact on the more generalised AI they are aiming for in fields in which there is not readily available human expert training data.

[shapenaji]

Doesn't AlphaGo answer that question? Train on KGS 1-9d game data (I'm still not clear if they used GoGoD or go4go with pro games as training data for Fan Hui or Lee Sedol version, there have been conflicting reports), play yourself and improve and then beat Lee Sedol. Deepmind have said they are also thinking about training a new bot from just the rules, to see if self-play without the initial guidance of human experts can reproduce master level skill (this is the approach they used on the Atari computer games playing AI); that would obviously take far longer but would also be interesting to see if it rediscovers all our standard opening theory, josekis etc or plays a unique style (wacky centre moyos?). That approach has deep impact on the more generalised AI they are aiming for in fields in which there is not readily available human expert training data.

[/quote]

Oh, certainly true, that's why I asked though, is it relevant what point it started off from? Do all neural networks converge to that point? or are there artifacts of starting at a weaker level that are maintained?

[cyndane]

Oh, certainly true, that's why I asked though, is it relevant what point it started off from? Do all neural networks converge to that point? or are there artifacts of starting at a weaker level that are maintained?

My bet is there are many minima to get caught in. It would be very surprising for everything to converge to the same point.