[chrisherring]

I've started playing Go against a computer (SmartGo iOS on a 9x9 board) but I find the hard part is determining where you're making mistakes, what is a better move, why the computer did what it did etc. For instance the computer may ignore a capture to play elsewhere and it can be hard to work out why. I learnt to play Chess when I was young so I can't make a fair comparison but it feels like Chess is much easier to self diagnose.

[panglott]

I took up Go about six months ago, after many false starts over the years. The secret for me was to stop playing computers and start playing humans online (I like online-go.com). Computers at lower difficulty settings make a combination of very strong moves and very weird moves, which are difficult for beginners to interpret. And the game is just a lot more fun when playing against people of similar skill level or a bit stronger, so you win about 50% of games and can understand your opponents' moves more easily.

[jonbarker]

The 12k bots on pandanet are good for learning how to make large groups live against a relatively good opponent, large group deaths are usually the most damaging thing to a beginner's results.

[gohrt]

Large group death meaning that the opponent invades before you have created eyes among your large-looking territory?

[NhanH]

In your large-looking territory, the territory would be the eyes.

Large group death just basically mean a large group of stones, it could either be an invasion group, an attack and run group that grows, or even the used-to-be-territory-but-now-not group

[chrisherring]

Sounds like good advice. I have noticed some of these weird moves using the hint function - at times they've caused me to lose badly so I've gone back and done something else and came out on top which made me question the reasoning behind it.

[mtgx]

I like this one:

https://www.flyordie.com/go/

[mrob]

One of the most important concepts in Go is the idea of initiative and forcing moves (see "sente" and "gote"). Making an unnecessary capture is giving your opponent initiative, so it's a "gote" play. They're free to play anywhere else on the board, which is bad for you.

Sometimes a group of stones is in a position such that no matter how their owner plays there is no way they can be saved (assuming correct play from the opponent). The stones are called "dead". Instead of capturing those stones, it's better to play elsewhere and retain the initiative ("sente"). You retain control of the game and your opponent is forced to react. You're not losing anything, because you can come back and capture the stones any time. If your opponent tries to save them you know a move that will prevent it.

Recognizing stones that can't possibly escape being captured, and knowing how to stop them escaping, is the study of "life and death". The better you get at this the less likely you are to waste moves or give away the initiative.

[tasuki]

This is the single biggest issue with introducing go to newcomers. In chess, the goal is crystal clear - capture the king. In go, the goal of having more points than your opponent is somehow very challenging to grasp in an intuitive way.

There's a common advice to lose your first 50 games as quickly as possible (http://senseis.xmp.net/?LoseYourFirst50GamesAsQuicklyAsPossi...). This will give you some basic intuition about the game. After a couple more hundred games, go becomes very beautiful.

[supercoder]

Thanks for the advice. Though I'm planning to be really good, so I'm going to lose my first 500 games.

[chrisherring]

Plus you can generally see who's in front in Chess just by counting the value of the captured pieces. Where as I would not have been able to workout the result of the AlphaGo v Lee Sedol games just by looking at the final game board.

The beauty of the game is appealing however so I want to get better at it. I'll working on losing more then :)

[tel]

Typically evaluation of a Go board runs something like

(1) Identify all the groups.

Most important are any groups larger than about 4-6 stones or groups that cannot be sacrificed for some other reason. You'll often see some straggler group which exists purely to disconnect your opponents groups keeping them weak. That one is small but important!

(2) Determine their status.

Essentially, their opponent decided it was time to kill this group and tried with all their might would the group be able to survive? You can be fuzzy about this ("there are a lot of possibilities for this group to live") or even economical ("it would take a lot of turns for B to kill W's group here, is it worth it?").

(3) Decide who "owns" every empty intersection.

Once you've identified the strong, living groups you'll see that they probably capture territory because they surround some empty intersections. These spaces can be invaded or reduced if they are too large (so that the opponent could possible attack inside and make a living group) or if they have open spaces (where an opponent could threaten to stream inside). There will also be a lot of points in the early and middle game for which nobody really owns and you can evaluate if one player merely holds greater or lesser sway over those intersections.

(4) Subtract maybe 20 points for each weak group that can't be sacrificed.

Saving groups that must be saved is an expensive operation and your opponent will probably cash in on you doing so.

(5) If it's early or mid-game consider each player's global influence.

For instance, a common trade is that one player will make a living group in the corner and take points while the other player will surround that group on the "outside" producing a large wall facing the center. This wall cannot claim to make territory since it doesn't completely control any location, but it certainly makes battles nearby tilt in the favor of the wall owner. This will translate to cash elsewhere for that player.

---

Don't try too hard at any step. Part of skilled judgement is being able to do each step accurately (and include more important factors) much in the same way that materiel balance in Chess can be hard to judge unless you know how to contextualize it.

Also note that a well-matched professional game will be decided by just 1 or 2 points. Until your judgement is very tight or the game is very near the end it can be hard to be precise enough to see that.

[kqr]

The general principle is that captures aren't actually worth a whole lot. Territory (area under your control) is where it's at. If a capture is a means of getting more territory, that's grand, but on it's own it's not worth much at all.

Does this help your self diagnosis? If the computer ignored a capture it's because it didn't actually lead to a lot of territory gained, whereas whatever move it played probably did.

[shultays]

While scoring go uses area you control + number of stones you capture right?

Somewhere I read about go it makes no sense to put stones on the area your opponent control because they will be captured. however

  wwwww
  w   w
  w b w
  w   w
  wwwww

in this case white has to put 4 stones (thus reduce area by 4) just to get 1 point. am I missing something

[mrob]

This is why I much prefer Chinese (and similar) rules with area scoring. In area scoring you get one point for each space you surround, and one point for each stone on the board. It's immediately obvious and simple.

In Japanese rules you lose points for playing in your own territory. The black stone in your example is dead, and a reasonable player would allow it to be removed, but there's nothing stopping them from claiming it to be alive and forcing white to play to capture it. At this point the only option is for white to dispute its life status, and to duplicate the game board and make a hypothetical sequence of plays to prove it to be dead without actually playing the moves. There's a lot of added complexity for no real benefit, because the end result is usually identical.

[Jach]

I still prefer Japanese counting since it takes a lot less time to count (at least for me) a full 19x19 board. The only way white loses points (relative to black) by being forced to "play it out" is if black passes -- but the AGA resolves that by requiring you to give up a pass stone as a prisoner when you pass, and if the game has ended and you have to play it out, passes aren't allowed.

The next AI challenge is to have the AI make an argument that will convince everyone to adopt the One True Most Elegant Ruleset for Go. (http://warp.povusers.org/go/RulesElegance/ is an interesting page, along with his proposal: http://warp.povusers.org/go/MyRulesProposal/)

[mrob]

Chinese counting is just as fast once you're used to it. Remember that you only need to count one player under Chinese rules. And the AGA rules are a nice hack to Japanese rules to make them always give the same result as Chinese rules and avoid trolling/unreasonable plays, but it's still more complicated to explain to beginners than Chinese rules.

[dfan]

Yes, I always teach novices with area scoring, because all of the "if I played it out, I'd kill you, but I don't have to actually play it out" stuff is incredibly confusing to someone learning the rules for the first time. It's much nicer to have no penalty (other than opportunity cost) for playing these sequences out.

The AGA system of pass stones, bringing area and territory scoring into alignment, is very nice, but I think it's still a bit abstract for someone who's just starting out.

[mfoy_]

But in practice there usually ends up being no difference in score under either ruleset, or by at most 1 point difference.

[gohrt]

A computer will correctly score that b as a dead stone, since it can't possibly make life, and white can kill it. That insurmountable threat makes the territory white's.

If a human tries to argue that b is alive, then either (a) that human is a novice, you can teach them, or (b) that human is a poor sport, and you shouldn't play with them.

Japanese counting is like Chinese counting, but assumes your opponent is not a jerk without explicitly handling cases when that's not true.

[Spellman]

If a group is dead, it will be automatically captured at the end of the game.

So, in this case, the black stone is auto-captured, and you get the territory within as well.

So, yes, playing white stones to capture is actually detrimental to yourself. But also imagine if black keeps playing. They would be giving away free captures since black can never create eyes. And once the space is almost filled white can play 1 stone to capture all of the black stones.

Edit: Another way to imagine it is if neither player could pass. While yes White would lose territory as it played stones, it would be compensated by Black playing stones that will be captured in the end. Thus it becomes a wash, and thus why dead groups are auto-captured.

[chrisherring]

That thinking does help, I did read that on one of the sites as well but in some cases the capture appears to mean the loss of territory itself in addition to the actual pieces and it's hard to weigh up the value of each territory.

[Jach]

Another reason the computer might ignore a direct capture is because the stones are already "dead", it doesn't need to waste a move to actually capture them unless you futilely try to save them. Solving tsumego might help you see those cases better when it's not as simple a case of leaving your stones in atari. But yeah, even if you can "save" them next move it might still ignore the capture if there's a bigger (in terms of gaining / securing territory) move on the board, especially if your save actually isn't making them alive yet.

I think scoring and score estimating is one of the hardest things to understand (for full games I still often use a $5 android app where I just upload a picture of the board game, help it convert to an SGF, and then it auto-scores it, though sometimes it's faster to just count by hand), especially since there's multiple ways of doing it. With simple territory scoring when the game ends the dead stones that haven't actually been captured yet still get removed and added to the prisoners pile and the prisoners fill in your other territory so it might feel extra painful to lose stones... With area/stone scoring you still remove the dead stones but prisoners don't (directly) matter for the score, since you're just counting stones on the board plus uncontested area. But if a game was X moves with black having X/2 stones but white only having (X/2 - 10) stones on the board, clearly (assuming X is even and no passes, or if passes then giving up pass stones) 10 white stones were captured at some point, so there's already that relative difference in score before counting anything else. With some qualifications, the results (and the relative scores themselves) from each counting system will be the same. Thinking more about your relative score instead of giving each player actual numbers might help; I feel slightly better if I lost by 10 than if I had a negative score, say -5 to 5...

[jdietrich]

As others have said, there's no substitute for playing against human beings.

When you've played a few proper games, I recommend joining a teaching ladder. Your games will be reviewed by stronger players. In exchange, you are expected to review the games of weaker players. Teaching can be just as valuable as learning, as it forces you to clarify concepts in your mind in order to explain them. Even if you're a high-kyu player, you can teach a beginner the basics.

https://gtl.xmp.net/ http://senseis.xmp.net/?KGSTeachingLadder

[fdej]

Clearly the next goal for AI is a program that can teach Go better than human teachers.

[blubb-fish]

that's essentially a pyramid system ...

[JeremyBanks]

Pyramid systems involve exponential (or at least superlinear) growth which was not implied by the parent post.

[niij]

Why was the top player there all along? In a financial "pyramid scheme" it is because money streams to the top. In a teaching ladder, what malicious incentive does the person at the top have to where leaving would cause the whole system to collapse? Most likely, the other players at the top would be happy to take on the role as top player in the league and would continue to self-teach to maintain their place.

[jdietrich]

Exactly. A teaching ladder is simply a cooperative effort to share skills. No money changes hands and there is no benefit to being at the top of the ladder other than kudos.

[blubb-fish]

was just kidding ... LOL

[blubb-fish]

well, they promise you to be teached - but (and here's the catch) you have to teach yourself - but then at some point the upmost teacher leaves with his knowledge and the whooooole ladder starts to crash - so obvious

[jdietrich]

A teaching ladder is a linear system, not a pyramid. I get a lesson, I give a lesson, one for one. Reciprocity is used purely to prevent the free rider problem. Participants freely barter their expertise rather than paying for lessons. This is a perfectly stable arrangement - the ladder in the first link has been running continuously since 1994.

The person at the top of the ladder is essentially performing a public service, because they don't get the benefit of learning from a more experienced player. If they leave then the ladder doesn't collapse, the second strongest player simply becomes the leader.

Ladders are commonly used as a competition format in many sports. They provide an excellent way of giving players a challenging but enjoyable game. Teaching ladders are simply an elaboration on this format.

[romaniv]

>For instance the computer may ignore a capture to play elsewhere and it can be hard to work out why.

I think the biggest insight I had about Go when I was learning to play it is that the game commences on the entire board at once. At the beginning I was over-concentrating on a single "active" region. In reality, abandoning some "battle" and adding stones to an "unrelated" region is a perfectly sensible strategy as long as you think it will give you territory in the long run.

[theseatoms]

I've had the same experience. It's easy to get drawn into petty skirmishes and ignore the larger play for territory.

A good rule of thumb = don't needlessly increase the margin of your victory for any particular part of the board

[gnidan]

Something that Michael Redmond 9P said a few times during the AlphaGo match was about how there comes a point in any sequence where the next move is of dramatically lower value than the previous.

[tveita]

I've had a similar problem. I would like to develop an intuition for the game by exploring different variations, "if I do this, the opponent does this, but if I do this, the opponent does this instead", but last I looked I couldn't find a client with a good interface for exploring the game tree that way with a computer opponent.

I know playing against humans is better, but a computer opponent has the patience to watch me try hundreds of moves that they know are dumb, or let me cheat to test "what if" scenarios.

A variation explorer with a "What would GNU Go do?" button feels like it would be good for this.

One thing that has helped me is Tsumego problems, which cover local life-and-death situations. They usually include responses by the opponent to help you see why a move is bad.

[gohrt]

My trouble with tsumego:

I can sit down and maybe solve a tsumego problem, OK. Sometimes I guess a bit (MCTS!) instead of working out the whole tree, developing that intuition.

But tsumego are labelled with difficulty and that they are winnable.

But when I play a game, how do I know how if a certain local position is winnable, and worth spending the minute(s) to find how to win it?

[NhanH]

There are two big purposes for Tsumego, one is to train your reading ability, the other is to train your intuition and knowledge.

For the former, you need to read out ALL variations, and not just one branch. For the latter, and your question, it's just simply a matter of solving lots and lots of tsumegos. In real game, people missed life and deaths situation all the time :-).

[dfan]

You might do better with problems that ask you to determine the status (alive/dead/seki/killable/ko/etc) of a group of stones.

[joseraul]

There are a few (free) teaching ladders out there. The basic idea is that you have your games reviewed by players a few stones stronger than you, e.g. you are ranked around 10k and the reviewer 4k. Reviewers are strong enough to spot the biggest weaknesses in your play, so that you can improve quickly.

When I used the ladder, 15 years ago, their problem was that they had too many reviewers and too few students, so it was nice as a student :) You should give it a try.

http://gtl.xmp.net/ http://senseis.xmp.net/?TeachingLadder

[Jach]

See if you can find someone to play a simple Capture Game with you on the 9x9. That is, the goal of each player is to capture say 3 of the other person's stones, then the game ends. This will help with some basic things like making sure you're keeping track of liberties, seeing how you can form two eyes (a stepping stone to forming territory) and be uncapturable, seeing which moves require a response to save a stone/group of stones, and seeing when an attempt to save a stone would be pointless (e.g. you're in a ladder).

[svarrall]

I've had the same issue. It would be helpful to have some kind of 'ranking' for a move after it's been played so you know if you've immediately made a mistake. Often after playing a losing game I'll undo moves and try some variations, but I don't know how far to go back. Was it my 50th move that was weak, or the 10th one... I know a moves strength isn't immediately obvious (as was apparent in the alphago games) so it's strength could be adjusted as the game progresses.

[Jach]

Blunders are fairly easy to notice in-game or afterward in review, but yeah it can be frustrating when you don't feel like you made a big mistake but still lost by a lot. If you have a better player look it over though they'll at least find a bad / inefficient move pretty much every few moves that "only" loses you maybe 3-5 points but the accumulation at the end is pretty big. Personally I like GnuGo for doing some post-game move analyses (I use the GoGui client) but I think I'd fare much better having a good player review or do a teaching game... Still, it's kind of fun to look at its suggested Top Moves, or try letting it finish a game (Final Score -- I think its Estimate Score is pretty terrible, though AlphaGo's would be nice to have on hand...) with itself after certain points to see if any are decisive for it. Another fun thing I did recently with a wild full game was have GnuGo annotate every move with the top moves it would have picked along with marking dead or critical groups. (The command is: `gnugo --level 15 --output-flags dv --replay both -l input-game.sgf -o annotated-output-game.sgf`)

What I really want is a kibitzing clock when playing other people: http://www.emptytriangle.com/archive/show/8

[weavie]

I am finding doing Go puzzles help here. I have an android app, I'm sure there are iOS equivalents. The puzzles only seem to tell you when you have made a wrong move and not why, but after spending some time thinking about it it usually comes clear.

[hyperpape]

There are a set of forums where you can get reviews from stronger players at lifein19x19.com/forum/. Downthread, people have mentioned the go teaching ladder, which is a good site, but you may get faster feedback on the forums.

[Gepsens]

Sounds like the hard part was the whole game ;)