I’m a new Go player and was attarcted to it from chess, by what I perceived as a less formulaic and broader game, more of a strategical battle than a tactical battle. I’ve been playing correspondence games in here, since typical live games seem to have very short time controls for a game of go and are more akin to blitz/ bullet in chess, and was left with some question, specifically pertaining to correspondence games:
Are correspondence games just a battle of libraries (much like modern correspondence chess) in which most people are effectively trotting out the contents of online joseki websites and of waltheri’s pattern search site?
Am I actually under a misapprehension and the role of formulas (like joseki) and pattern libraries is actually far more important than it seems at first? It seems to me that if one player has access to an online library, then if two players of the same level are playing the one not using them is under a clear handicap.
Because you have such extent of information available online, and even engine analysis of positions, isn’t correspondence go really solely about best usage of said resources?
Again, all this is very similar to correspondence chess, I’m not shocked by it, nor really surprised, but obviously I want to understand if I’m bringing a toothpick to fight a nuclear war.
Correspondence players will look up josekis and openings, for sure, but this gets you hardly anywhere in practice.
Nope, because most the subtleties of most joseki escape your average correspondence player anyhow.
There’s not a whole lot of difference between playing solid moves and playing something that is “joseki”, in terms of outcomes of a DDK and even SDK game. There is so much left of the game after those sequences are played.
Nope, I don’t think so. Players who are ranked higher win more against players who are ranked lower in correspondence, very much in line with online rankings. This is because they know how to play better, not because they are better users of joseki dictionaries and waltheris.
Luckily, the ammount of possible moves is so high, even waltheris database really will not get you far Just try it on some of your finished games or just with hypothetical play. I bet you will run out of Waltheri moves before you reach move 20.
And yes, if somebody wants to cheat they can always just boot up Leela and relay moves, but actually I think it is even less common in correspondence games, as such a cheater usually realises only relaying someone elses moves for three weeks is rather silly and not fun
Welcome, fellow chess player. I think you’ll find that pattern libraries are actually far less important than it seems. As you said go is a less formulaic and broader game. Getting some moves from a library won’t be that much of a help because the game branches out into a completely unique position very fast. Looking up joseki on the fly helps to not get destroyed by a joseki trick but that’s about it.
I think these resources don’t affect the game results that much. Using your brain and thinking about what you play has far greater effect still lots of players don’t do that and play by the feel.
In my very humble opinion, the mathematical/ programmer (?) approach that is used to understand/ explain the possible outcomes of the game might create the misconception that Go isn’t a very creative, personal game.
Cooking and especially baking are basically chemistry, but I doubt a Nobel prize winner chemist can beat a chef just by having access to every chemical reaction possible.
If you find the time settings of open challenges too fast, I’d recommend creating a “custom challenge” with more fitting time settings. (Avoid “simple” time settings, they have many bugs).
The OpenStudyGroup (OSR) is a place you can find players to play games on a slow pass (30m+overtime on 19x19), and the possibility to get a review or some general tips afterwards.
At beginner level, your opponent won’t have a big advantage by trying to play joseki sequences, as you probably won’t follow up with a move that’s in the joseki library. Your opponent then is left on his own.
There are a hand full of short sequences, that are helpful to know after you learned the basics of the game, but at DDK (double digit Kyu) level, the use of Joseki libraries is very uncommon. (More than half of the OGS player base is DDK by the way).
Until a few months ago I used joseki dictionaries and the analysis tool in my correspondence games. One day I decided to stop that. My rank was at about 6k at that time. Today it is at 6k. Now, there is a time lag of course and some games that I am still playing have started before my switch. So I’m not saying there is no impact at all. But I am pretty sure that stopping to use both, joseki dictionaries AND the analysis, will not impact my rank by more than 2 stones, more likely just one. So I would guess joseki dictionaries alone are probably worth about half a stone for me. Far less than my usual ranking fluctuations.
It may depend on the playing strength of course. When I looked up stuff in a dictionary, I sometimes understood the moves and in most cases at least tried to pick moves that I thought I understood to some degree. Still I was often lead into positions that I did not really understand. A 25k will most likely gain absolutely nothing from joseki dictionaries, as they are likely to get it always wrong as soon as the opponent deviates. (Unlike me, who only gets it wrong in 90% of the cases.) A 5 dan however may actually correctly judge which of the many correct josekis to use and how to play on. So on that level, it might be entirely different. Then again: how would I know?
Yeah, joseki databases like josekipedia.com are just that: databases. They do contain a lot of information, but they wont tell you which joseki is good in your game with your overall board structure and overall sitaution.
In josekipedia, even with the simple “black 4-3, white approaches low on 3-5” has total of 86 899 possible continuations. Its not very useful to know all of them (no-one does) or knowing how do they play out, but rather, the trick is to know what you want to achieve in your game with your overall board situation.
And indeed, games are not decided on josekis, they help you little bit in the opening but its not really that significant.
About joseki and using pro game databases: unless you are aware why a joseki is good and why pro players play certain move, you’re not going to be able to take advantage of your opponent when they stray apart from conventional moves.
Actually, many joseki are not even the optimal play, as we have seen since AlphaGo taught us improvements of century-old patterns. It is likely that current day AI also makes mistakes, and that there exist even better joseki out there.
And, how good a joseki is, will strongly depend on the overall position of the game. Some joseki are considered good, but also very difficult to use.
Finally, as others have said, joseki and pro game databases will only get you about 20 moves into a game that lasts 300. I’d say that most points actually get distributed in the midgame, rather than the opening. There are even people who can gain a significant advantage from being good at just endgame.
The only real mathematical approach to go strategy that I’m aware of is endgame analysis (and only the very last part of endgame). The opening, midgame and early endgame are all done more intuitively than mathematically (not that those two have to exclude each other, on the contrary, mathematics done well should be intuitive).
It may need a database…but hard to not call modern AI a mathematical/programmer approach…
and it did turn out to be strongest of all. So perhaps that idea is a bit lacking
however, more important to human level play… WRT to joseki, they are just optimized corner patterns… from both players point of view. There is not like “lower right corner joseki”, because there is not a best pattern for the lower right in a vacuum. Each joseki option will accomplish something different. If they are just generically applying joseki, OK they will probably not die… I don’t know that the outcome will be so amazing or as powerful as you may be expecting if the player was not considering the preferable outcome when selecting what joseki to use.
It’s actually not that hard. We have literally no idea how modern AI think. They learn the stuff themselves by practice. We don’t teach them strategies, we don’t tell them how to count or how to read, we just let them try and as if by magic they learn the game. If I would describe how an AI knows what to play, it is not because it has been programmed to play a certain way, but it is because of intuition gained from practice.
It’s like Gia’s comparison from before: baking bread is not much more than a chemical process, but a good baker is not a chemist. Strong AI are made with programming, but they know how to play without involving programming.
Mmmmm, I think it is pretty easy to say there is nothing going on in the box but math and programming.
It is not thinking at all, its performing the steps and pulling data from databases.
You didn’t write the steps, and don’t know them. The steps are not straight forward or something that is going to be reduced to 10 bullet points you can study…
But there is just no way to argue it is not math and programming lol. Sorry man.
I’m not saying it’s not based upon math and programming, I’m saying that building a modern AI to play Go is not a mathematical / programmer approach to playing Go. Our brains function through electrical impulses flowing between neurons, but that doesn’t make an electrician an expert in psychology (nor in programming).
There’s a difference between the two things. An AI is a computer program (duh), but the way it plays go has nothing to do with programming.
As for what I would actually consider a mathematical approach to Go, it is more along the lines of this type of thinking
Greetings! Let me re-phrase your question in a couple of different ways that may shed some light on things.
So, given the assumptions that
there are many high-quality AI Go tools out there, and
the main priority for online Go players is to increase their rankings by any means available (even if that means cheating)
the obvious question becomes, “Well, why isn’t everyone using AI assistance and cheating all the time?”
The surprising answer - as several people have already pointed out - is that it is not as much FUN as ranking up due to your own time, effort, and skill growth.
Now, do people cheat and use AI assistance on OGS? Yes, they sometimes do - and the OGS admins do an incredible job in identifying those situations and handing out consequences as appropriate.
However, in my opinion, this seems to be the exception rather than the norm, and the main reason is that rather than playing a game and learning from it, using an AI’s moves turns the active player into a passive transcriber - reduced to doing data entry without even necessarily understanding WHY the AI’s move was better than the moves they could come up with on their own. It turns out that - for most players the priority isn’t their abstract ranking - it’s their actual skill in being able to play the game, and the joy that comes from real growth and improvement.
As a counter-point, there have been many interesting threads on the OGS forum regarding different ways to use those AI Go programs as learning tools without resorting to cheating - ways that actually increase one’s skill and understanding of the game over time:
Cooking and especially baking are basically chemistry, but I doubt a Nobel prize winner chemist can beat a chef just by having access to every chemical reaction possible
Lets look at this…
Cooking = go
chef = human acting on intuition instead of pure science
chemist = programmer, just using base rules to produce a result without any intuition.
We did this one, the chemist (deep mind) beat the ever living hell out of the chef (Lee)
Good movie
I’m starting to repeat myself, so this will be the last reply about this:
Programming an AI to do the thinking for you is not the same as solving a problem with a programming approach. If the AI was playing go following some (humanly interpretable) program, we would be able to understand the rules it uses to decide on moves. “Why did the AI invade the 3-3?” cannot be reduced to some rules, the only thing we can say is that somehow the neural network gave that move the highest score.
Compare it to neural networks that recognise handwriting (kind of the archetypical AI example). A programmatic approach would involve describing the structure of the shape (if it has two holes, then it’s a “B” or an “8”, if it has one hole then it is an “o”, “p”, “q”, etc. if it has no holes then it is an “x” “z” “c”, etc.) and making some stepwise heuristic of determining which symbol you’re seeing. This is not at all what neural-based AI do, though. The patterns that AI look for are very different from anything humans can work with. See for example this excerpt of a 3blue1brown video on neural networks, where he shows what a neural network actually does to recognise handwriting (the whole series is recommended, by the way). It also shows a major flaw: his neural network will recognise certain random noise as if it is clearly a certain letter. There are even ways to fool an image recognition into giving a nonsensical answer by changing just a single pixel.
If you ask the programmer to make some changes to his code, so that the AI will stop invading the 3-3, for example, then the programmer has no way to implement this: the only option is to start from scratch and give the program a penalty for playing 3-3 invasions before you start training it. If instead the AI was using a mathematical heuristic to play Go, this would not be a problem at all, since you could just change the heuristic slightly.
Buildings are made of stone, but that doesn’t make a geologist good at architecture.
Sound is just physical movement of air, but that doesn’t make a physicist a good musician.
The strongest go players are built by programmers, but that doesn’t make a programmer good at Go.
I’ve ordered it yesterday, together with John Conway’s “On Numbers And Games”, but it takes a few weeks to arrive…
But this is indeed how Go would be approached mathematically. Very distinct from anything any go player would use, and only useful for those who want to play the last 20 moves of the game flawlessly. For the rest of the game it has no use.
Just try it on some of your finished games or just with hypothetical play. I bet you will run out of Waltheri moves before you reach move 20.
