![]() StartingPoints Referenced by
|
etrynus
PageType: HomePage
Hello all. My name is John. I am etrynus on KGS. I'm not much for writing these things, but since it's go, so I'll make an exception. I seem to have a lot of thoughts about go... Most recent change 5/13/04: [432] How to Start Memorizing Games
[1]To begin, I'm currently a freshman at MIT, where I am enjoying the broadband. It's probably the biggest reason why, for the past couple of months, I have gotten quite addicted to online go. Already, I think it's going to be something that will consume my entire life. The fact is, I just REALLY like go. When I tell people that, they go 'ah, go, interesting', but unless they are addicts themselves I know they don't understand. It's probably been said many times before, but it's hard to find such a game that is so complex yet on the surface pretty simple. Go is a perfect information game, yet we are all so far from perfect. Yes, don't deny it, even the pros suck. It's hard to believe someone said they could take 3-4 stones from God - I'd say the top pro would need at maybe even nine. If you think about it, on every move there is a set of moves that lead to the optimal result for the player. I'd suppose that the number of moves in this set would generally be around one or two (maybe more in the fuseki). If you consider the number of moves that are considered 'playable' by a pro, then it would seem that only a small percentage of the time does the correct move on the board get played. It could be 5%, or 30%, who knows. Either way, even if each 'mistake' constitutes on average as a point below the maximum, I think a pro could easily lose to the perfect player by about 50-100 points. Maybe 9 handi, maybe 6, but definitely more than 3-4. Anyways, in line with the complexity, to me the other fascinating aspect of go is in the mysterious learning process. It's difficult to describe how you've improved, and near impossible to put in words. Sometimes, when I'm looking back at games I've played months ago, I actually pity myself for all the stupid mistakes I made. It's easy to feel content winning a game I've just played, but when I look back at games I won while weaker, somehow it feels cheapened. And the cycle always continues... Anyways, in regards to learning go, I'm reminded of how society always feels it's at its technological peak ("everything that could be invented has been invented") - at any given moment I always feel that there's no way my improvement can continue, yet (luckily), it does. [2] Ranks [21]Ranks are great. If it weren't for ranks, I'd be frustrated as hell. That's not to say, however, that I'm in the 'win at all costs' crowd. I'm fine with winning and losing, the main concern is more with improving my play. Thus, my personal policy is to always make the play that I think is the best on any given move. Even if the game hinges on killing my opponent's group, and my opponent makes a move which I read to be death in gote, I'll gladly play elsewhere to fill in a half-point ko. It leads to embarassing situations once in a while, but I think every move should be the one that you think is optimal. Unless it's a tournament or something, it's a bad habit not to play that way. (I define the best play as the one that will maximize the point differential between you and your opponent given perfect play for the remainder of the game.) As for myself, I am currently a 9k playing almost exclusively on KGS. I learned about the game of Go back in Aug 03, but aside from downloading Igowin I didn't see the light until Jan 7, 04, when I stumbled upon KGS. This is when I'd say I really began learning Go. At that time I started out at 20k and since I've kept track of my progress in my KGS profile, which is right below. It relates rank to the number of games I've played, including unranked and teaching games.
You might wonder why I even bother to keep track of all this stuff. Well, the initial reason was that when I had completed around 10 games, I ran across someone's profile proclaiming '1000 games to Shodan', with a list similar to this one. So I thought it was a great idea, and put one up of my own. After about a hundred games, I got sick of keeping track, but continued to do it for a new reason. Not only do I want to reach Shodan as fast as possible, I would like to do it playing as few games as possible. I see people who play a bucketload of games but don't get very far - my purpose is to take as much as I can from every game, no matter the result... winning by 50 just means someone stronger could have won by 150! Anyways, my personal goal for reaching shodan, taking into account my progress thus far, and the inherent logarithmic rate of improvement in go, is between 1 and 2 years, and between 1000 and 2000 games. Anything less would be nice, anything more would be disappointing. KGS Teaching Ladder [22]This teaching ladder was recently formed in KGS, and I am in complete support of it's development. I'm sure there have been numerous attempts in the past to get something like this going, but this one looks to be here to stay, especially thanks to DrStraw for posting a reward to motivate teachers to post more teaching games. I post games every so often but also don't think it rude to message me for a teaching game (also don't think it rude if I decline). Game Ideas [3]There's a lot of stuff you can do besides play the regular 19x19 game to improve in go. In fact, I think that periodically deviating from the normal version could actually help you improve faster. I would like to try some of these ideas out in some teaching games but I'm afraid that the student will be unsatisfied with the result, so either I'll make it in the game header that something weird is going on, or, if anyone would like to try something out, message me. Shinogi Master [31] I hope I'm not using Shinogi wrong. I just wanted to sound cool... umm.. Anyways, the basics of this game is that white has a complete square enclosure, and black tries to live in it. Since I've never tried it, I would say that white should start at a 9x9 empty square (which I think would be pretty impossible to live in), and if black fails he tries again at 10x10, then at 11x11, and on until he succeeds. Then, the players reverse, and white tries to live inside black. What I am REALLY interested in, is how the difference between the size of the boards in which black can live and white can live correlates to the rank difference between the players. I'm sure it wouldn't be linear, and there would be a huge mass of people at a certain size (maybe 13x13?), but it would still be interesting. Fighting Game [32] This is a really simple idea, which I've tried once and ended up liking. Set up an unranked game, and be as agressive as possible, without worrying about the points. Instead of extend, pincer. What you want to end up with after the fuseki is about six groups on either side, all pretty weak. Just always remember to cut and it should happen. The amusing thing is that you don't even need to tell your opponent what your intentions are :-). Thus, instead of worrying about making the correct play, you can focus on local situations around the board, which should improve your reading quicker than normal games would. I think the trick to this is that you should not play it so much that the bad habits become a part of your game, but enough so that you get what you want out of playing it. Undos Allowed [33] It sucks to lose a game due to a stupid blunder, and your opponent won't let you undo the move. While it may not seem like it, it's also bad to win a game due to blunders from your opponent. If you are actually better than someone, you should be able to win, not because of the mental lapses, but because you have a better understanding of the game than he does. In other words, ideally, after you win, your opponent should have no concrete idea as to why he lost. I have the unsubstantiated opinion that we are approximately 5 or 6 ranks stronger than our normal level of play. Meaning that, as a 9k, I have all the mental resources to match a 3k, it's just some silly mistakes together with the more subtle ones that make me a 9k. Similarly the 3k would have the capacity to match a 3d. I think that's why I can easily understand games of low kyus, but when I watch high-dan games, I seldom know what is going on. I could think and think and think, but pitted against a high dan, I simply don't have the knowledge to match him. (Obviously the 5-6 stone difference shrinks as we go higher up the ranks, a 1d for instance probably does not have the capacity to match a 7d). Anyways, whenever I get the time and the opponent, I want to play a game where undos are freely allowed, even after several (or tens of) moves have passed, and it is discovered that the line of play is not favorable towards one side or another. Ideally, the completed game would be at least a thousand moves later!, when both players are completely satisfied that each move is to the best of his ability. I think then, when a stronger plays goes to review it, each comment he makes will have significant meaning. Go Teaching Ladder Variation [34] Here is an idea. While you are playing a game with a sizeable time limit, on EVERY move, take about ten seconds to write down what is going through your mind when you are playing that move. This could be anything from 'connect two weak groups' to 'my gut tells me this works' to 'reading sequence' (which means that it's purely tactical and you can't put it down it words easily). I think weaker players would benefit from going through your game, to learn what is it that goes through a stronger player's mind when they play a game. I'd expect this to be more valuable near the fuseki and middle-game, and less towards the end. I'll try to post up 'what a 9k thinks' sometime. How Many Stones [35] How many stones does it take to claim the whole board? Basically, I could possibly put 25 stones on a board such that another 9k could not make life on the board. In contrast, I would probably need about 50 stones to prevent a pro from making life. More on this in a little bit. The game part of this functions a lot like Shinogi Master, where I want to see how the difference in actual rank between two players correlates to the difference between the number of stones needed to prevent the other from making life. This also brings me to the more general topic of... Musings [4]How Many Stones (continued) [41] Say that I am a perfect player, playing against another perfect player. Then, there is a number of stones (n), under which my opponent cannot make life, even after perfect play from both sides. Thus, with n moves, I have claimed 361 points worth of territory under area scoring. I think that if we can estimate n, we can make a number of interesting (though far from scientific) assumptions about go. For one, say that I only made n-1 moves, meaning that my opponent COULD make life in my enclosure. That last move is then worth a LOT of points, definitely more than 361/n. However, if I made n-2 moves, I could probably just remove a stone near the one where my opponent would have made life anyways, and thus that move should not be worth nearly as much points. What does this all mean? Actually I'm still compiling my thoughts. In the meantime I'll try to experiment around the value of n. Perfect Play [42] Back to the topic of how many points a pro is behind perfect play. First, it's almost impossible to tell how many points throughout a game that a player will lose, since both players are continuously making mistakes. For example, take a situation where the life of a major group will depend on who has sente, but both players are under the mistaken impression that the group is already alive. Thus, when a player tenukis, his play loses about 50+ points from the perfect. Then, if the two players continue like this for ten moves each, you couldn't say that each player is playing 500 points away from perfection. Thus, perhaps I was wrong to say that a pro was not 3-4 stones from perfection (PERHAPS, but I still think I'm right) - even if a pro does lose 50-100 points over the course of a game, that doesn't mean that he would lose by that large of a margin playing against the perfect player directly. Anyways, how then, could we directly calculate how much a move is worth? Well, this is another unfinished idea of mine (hence musings). Some more side questions pertaining to the optimal first move... Does the perfect black open on Tengen? I'm confused at how people can be sure that it is. Just because tengen is unique, does not mean it must be the best play. At any point in the game, there can be a number of best plays that are all worth the same amount of points. Perhaps there are even hundreds of variation of the perfect game which all end up with the optimal result for both sides. Anyways, first, we know that tengen is the best opening for, say, a 5x5 grid. We also know (I think) that tengen is not the best place for a 101x101 grid. From this, I think that there is a strong likelihood that the relative value of tengen starts high on small boards and steadily decreases as the board size increases. At some point between 5 and 101 this value will cross a 'line', thus making tengen the best play for all boards smaller, and not the best play for all boards larger. Is there an nth line on which the perfect play is on for EVERY board size? I'll try to clarify this question a bit. Say that we eventually find out that 11x11 is the largest board size where tengen is the best first move. In that case, the best opener is on the SIXTH line. What I'm asking is that given that fact, is the perfect opening for all board sizes greater than 11x11 also on the sixth line? The feeling I get deep down inside is that it is. Improving at Go [43] Blind Go Simplified [431] I think about go a lot. It is the nature of being an addict. However, since I'm not that good at go, it's difficult for me to run too many variations in my head and actually benefit from my thoughts. Recently, I've discovered something that is working for me when I am not around a go board, and might work for anyone else at any level. You start (in your head!) on a 2x2 board. Read through all the variations, and where do you end up? Pretty easy right? Now do it on a 3x3 board. Anyways, the point is to take it as far as you can from there, depending on your level of course. A tip for if things get a bit confusing, is to set in stone the first few moves of the sequence, since it's hard to figure out what the best moves are in the beginning. For example, on a 3x3, you first think, black opens in the middle and white answers in the corner - then read out all the variations. Then you think, black opens on tengen and white answers on a side - again, read it all out. This definitely helps on 4x4. How to Start Memorizing Games [432] A solid piece of advice given by many stronger players is to memorize professional games. Having reached the single-digit kyus, I've recently started to feel obligated to take up this way of improvement. If you've ever tried memorizing a complete Go game (or remember when you first started), it's not that easy. I consider myself a pretty decent memorizer, especially with pointless things like long strings of numbers and such, but when I finally sat myself down with a Takemiya game, it wasn't as I expected it to be. Not only did it take me quite a while to get the whole thing in my head, but coming back to it after a few days, I would mess up the order of sequences, especially throughout the middle-game. The moves within sequences were mainly intact, however, leading me to believe that my real problem was that I didn't quite 'get' the overall flow of the game, thus jumbling sequences of moves, and not specfic moves themselves. Saying that, here is my suggestion of how to gradually increase your ability to memorize games, by starting with your own. The key is in training yourself to subconsciously make a mental note of your thoughts during every move you make. Done with the KGS client (which was perfect for this idea):
The long-term goal of this exercise, as mentioned earlier, is just to get your subconscious into memorizing your games with minimal effort. It's common to hear about how pros can easily play out games they've just played - once you can perfect this exercise, you will be like that too! Of course, you hear about pros who can replay games from years ago, but uh, don't worry about that just yet... The final thing is to extend the skill of memorizing games you've played to memorizing pro games. If you've mastered the former, the latter should be pretty easy. The big thing is probably to play out the pro game as if it were you who was playing, and hopefully your subconcious will do the rest. Note: I've done this once so far (I only thought of this idea today). In a 226 move game, I made errors on 31 of my opponent's move, and 18 of my own! Man, I've a long way to go... My only excuse is that I thought of the idea AFTER the game was played, so that during the game, I wasn't trying to memorize anything, just play regularly. From now on I'll do it conciously. Computer Go [5]I don't think I (and like-minded people) could seriously get into go without aspiring to write a Go-playing program. Yes, the best bloody Go program ever. However, I'm still new to go, and on top of that still quite new to any real programming, so I don't really expect to do anything right now. I'll try to get started in a couple of months (it helps that I'll be taking some more computer classes in the meantime). Before that, however, there are a couple of things that I think I COULD do now. This would get me warmed up for an actual program. So far I've only thought of one 'warm-up' activity. [51] The Most Important Point [511] I'm positive this has been done before. You take a program with all the rules of go written in. Every turn, the computer simply looks at the lists of legal moves, and plays a random one. To avoid an indefinite game due to repeated suicides, the set of legal moves would have to be narrowed down such that a color would not play in it's own one-space eye. That is, Black would not play on a point where it's four sides were black, and at least three of it's corners were black (thus making it not a false eye). The program would then play itself a billion times and we would get some sort of distribution over which points are most important. Obviously, there are a few major flaws with this plan. One, it would take about 10,000 games to verify within standard deviation that, say, the 1-1 point was 'only' involved in 49.5% of the victories, or something just as unuseful. Second, playing randomly is a kind of sketchy idea for statistical analysis, if you stop and think about it, which I won't do :-) Actual Plans [52] In my opinion, the problem with go programs right now is that the people writing them don't know everything about go. Even if a pro was a programmer and decided to sit down and write something, the program's strategy would still be inferior to that of the writer. Therefore, I think the VERY BEST go program that is able to be written using current methods (ie joseki studying, pro games studying) will only be a program that has adequate strategy (albeit spectacular tactics). That would end up pretty darn good, but to move from there to perfection would almost be impossible. This is especially revelant to myself because if I followed the same path, I would always be shadowed by better and more experienced programmers (not to mention better and more experienced go players), and while I might end up producing something, it'd be pointless. Thus, I have to somehow 'cheat', and figure out a different method of writing a go program. I think, in the end, the best go program will have no knowledge of human concepts such as shape, kikashi, sente, etc. The program will just play. If you create the perfect score estimator, you have created the perfect go program. That should be pretty obvious. However, if we have a mediocre score estimator, can we turn it into a good go program? Maybe... First, the basic program would be simple - on the outmost level, the program would evaluate a number for each move, and play the best number. To evaluate the number, the program could use any number of score estimators, such as the KGS one. Obviously, it'd play at about the 30k level, but there are many possible enhancements. Look-ahead. Pretty much every board-game playing program has look-ahead. Instead of picking the move your score estimator determines as maximum for the next move, the program picks the move which the score estimator determines as maximum after a number of moves, provided the opponent plays like the program itself. Recursive look-ahead. Going back to the base program, say that you ask the program to play out the remainder of the game given what it calculates to be the best moves. At the end, it takes that result and replaces that as the value of the first move it played. I can hardly understand that sentence myself, so here is an example. Suppose it is move 50, and you are asking your program to play move 51. For each of the legal moves remaining, the program will play out the entire game using what it thinks will happen. So if the move in question is R6, and at the end of the simulation the score turns out to be B+5.5, then the value of R6 will be replaced with 5.5. After the program has done that for each move, you have, in effect, created a NEW score estimator! This will probably take a huge amount of computation though. The even bigger possible problem is the question of whether the recursion will improve the score estimator, or just make it's flaws more evident. Will have to try it out to answer that. Extra score estimators. Simply, using a variety of score estimators and implementing the techiniques described to get a better evaluation. Breadth/depth of look-ahead. One common method of cutting down the run-time is by taking out moves that probably won't work. In the case of the recursive look-ahead, for example, I would take out all but the best ten moves after the first recursion, and work from there. The danger in this is that then the program may not 'see' tesuji which at first appear to be detrimental to the player (sacrifice/nakade tesuji). Anyways, perhaps the breadth/depth of the search can also be dependent on the game board. For instance, in the beginning, since breadth is more important than depth (at least I think it is), looking at 10 variations 3 moves deep (1000 possibilities) would be good, while in the mid-game and towards the end the program could look at fewer variations deeper while keeping approximately the same number of possibilities. My own score estimator. This is perhaps my most unfinished thought. I keep thinking that somehow someway I can ask a program playing against itself randomly to make a good score estimator. Here is a bit of what I am talking about... In it's most general form, the value of each empty intersection is a function of every other intersection on the board. The function would most certainly be a nasty one, but maybe it is possible to simplify things. I'm going to present the most simple case I can think of as an example, and say that what I eventually want to do is like this, except a lot more complex. So on to trying to approximate the 'perfect function'. First, I'll define U(x,y) to be a function of the point (x,y), which will return 1 if black, -1 if white, and 0 if empty. Then my grand function could possibly look like... F(x,y) = (a_11*U(1,1))^b_11 + (a_12*U(1,2))^b_12 + ....... (a_1919*U(19,19))^b_1919. a_** and b_** are assumed to be constants. To figure out the value of these constants, that is where the random self-go player comes in. It plays games where it holds most of the values of a_** and b_** constant, allowing a few to vary, to try to figure out what is the optimal value for each of the 361*2 constants for each point. In the end, the program will have a value for 361^3 constants (which isn't that big), and that will be the score estimator. How effective will this be? I'm pretty sure not very, but it is not that hard to do, and can be easily improved by experimenting with better forms of the general function above. End [6] I think that's all I have to say for now. It's really really lengthy already. I actually started writing this (in note form) for myself to collect all my thoughts on what I wanted to do, and then decided to put it into paragraph form for SL. I don't expect too many people to read it, but hopefully at least some other people can give me some second opinions on my ideas. I know absolutely nothing about wiki, if you haven't noticed. Feel free to make this page look better. I think I'm going to stop keeping track of how many games I've played after all. In the end, sometimes I want to play some games for fun. I also want to play blitz games without feeling bad about getting closer to my 1000 games limit. So screw the limit, my only goal for shodan now is time. On the same note, I've created a new account yoshiyoshi on KGS solely for playing blitz games. It is currently on 6k, which is nice, but the only reason is that I win games based on poor mistakes from the opponent, not because of actual skill. For some reason I'm pretty good at keeping my head during blitz. It's fortunate since I can play stronger players in unhandicapped games without doing anything sneaky. This summer I'm going to be without internet access so somehow I need to keep on improving without playing everyday on KGS. I've found a couple of PBEM games, but any suggestions on what to do/study would be much appreciated. 3 months wihtout online go... :-( This is a copy of the living page "etrynus" at Sensei's Library. ![]() |