I tried playing an opening (up to move 46) against the AI from the Play Go Against a DCNN page.
My mistakes totaled 17.3 points (and the AI's mistakes totaled 17.4 points). Actually, it was mostly the same mistake, as we both kept ignoring the most urgent point on the board.
I am not sure if I will bother to post a review, but I think it is worth noting that just because I have reduced my point loss against one AI does not mean I have improved by a similar amount in general.
At some point I'd like to practice more openings against the linked AI or something similar, i.e. a policy model trained to emulate human players. I'd be interested in alternatives to the site I linked, because it doesn't provide a way to download an SGF, so I'm left to transcribe it by hand.
Also, this reminds me that I wanted to talk about metrics a bit. I've been judging my performance purely in terms of point loss, but I think this may have drawbacks. For instance, it may lead me to avoid difficult variations, even if they would be equally difficult for my opponent.
I am considering a new metric where I would subtract half my opponent's score loss from my score loss, so e.g. if my point loss is 20 points and my opponent's is 6, then my adjusted point loss would be 17 points. This would help encourage me to play variations that are difficult for my opponent, and also penalize me less when both players repeatedly fail to notice something.
I won't be adopting this immediately, however, as I still need to finish my current kifu study project, and changing metrics mid-way through would just be confusing.
Opening study with KataGo
-
hakuseki
- Dies with sente
- Posts: 100
- Joined: Thu Oct 29, 2020 6:18 am
- Rank: KGS 2 dan
- GD Posts: 0
- KGS: hakuseki
- Been thanked: 15 times
Technique notes for games 47-56
I'm about halfway through writing up these notes, but I feel like it is taking too long and I'd like to move on to playing more games, so I may not finish.
Anyway, while playing my next set of openings I'd like to focus on the following questions:
Am I moving in the right direction? / Am I approaching from the right direction?
Am I using the right technique?
Here are my incomplete technique notes:
Technique notes for games 47-56
(Note: these games were played with group tax, which affects opening strategy somewhat)
Surrounding is better than approaching In this position I played. My reasoning was as follows: Black approached, and white played a fourth-line extension (see marked stones). A double approach is good in that situation. The extra stones don't matter. In fact, the extra stones do matter, because black has a chance to surround instead of merely approaching. Black should have played at a instead. As it is, invites white to answer at a, keeping black separated.
Hane a contacting stone I kind of think I should have addressed this in my joseki review, but I'm including it here anyway. I played
but I should have played at a. Perhaps a hane against a contacting stone is intuitive, and is weak for black as well as white.
Actually I don't think this is so clear-cut, which is why I now hesitate to discuss this as a "technique" mistake. If white responds at b, as happened in the game, then white has been tricked and the sequence is not bad for black. Instead, white should block at c, but I think this requires deep reading to verify.
Two-space extension to create a base I am not sure why I played only a one-space extension here, but it should clearly be a two-space extension, both for stability and to pincer white's marked stone.
Capture or connect? White plays the atari at
. In this case black should connect at a. In this case I think a does more to develop the center; although white's stone is not captured immediately, the remaining aji doesn't seem very useful for white. Still, I played the wrong move of b here, losing 2.4 points.
I think this kind of capture vs. connect situation is interesting, however, because this decision can go either way depending on the surrounding context. I'd like to collect a set of example positions like this to compare these moves. Once I have several such examples I may create a new post.
Black is pushing from behind. How should white respond?
The answer is the one-space jump at a. This move is faster than a nobi, but it's important to check that the weakness of the one-space jump is not exploitable:
Running technique Solution follows:
I've previously discussed how this is a nice shape. The 3-3 invasion at a is also a good option. My move was the one-space jump at b, but this is inferior by about 0.4 points.
Actually every white move here is a mistake.
should perhaps be at a. But I will especially remember that 
is aji keshi. Black will connect against the peep, which removes the possibility of a later atari by white stones at b and c. Instead, the kosumi-tsuke at d is a tesuji to make shape.
Anyway, while playing my next set of openings I'd like to focus on the following questions:
Am I moving in the right direction? / Am I approaching from the right direction? Am I using the right technique?Here are my incomplete technique notes:
Technique notes for games 47-56
(Note: these games were played with group tax, which affects opening strategy somewhat)
Surrounding is better than approaching In this position I played
. My reasoning was as follows: Black approached, and white played a fourth-line extension (see marked stones). A double approach is good in that situation. The extra stones don't matter. In fact, the extra stones do matter, because black has a chance to surround instead of merely approaching. Black should have played at a instead. As it is, invites white to answer at a, keeping black separated.Hane a contacting stone I kind of think I should have addressed this in my joseki review, but I'm including it here anyway. I played
but I should have played at a. Perhaps a hane against a contacting stone is intuitive, and is weak for black as well as white.Actually I don't think this is so clear-cut, which is why I now hesitate to discuss this as a "technique" mistake. If white responds at b, as happened in the game, then white has been tricked and the sequence is not bad for black. Instead, white should block at c, but I think this requires deep reading to verify.
Two-space extension to create a base I am not sure why I played only a one-space extension here, but it should clearly be a two-space extension, both for stability and to pincer white's marked stone.
Capture or connect? White plays the atari at
. In this case black should connect at a. In this case I think a does more to develop the center; although white's stone is not captured immediately, the remaining aji doesn't seem very useful for white. Still, I played the wrong move of b here, losing 2.4 points.I think this kind of capture vs. connect situation is interesting, however, because this decision can go either way depending on the surrounding context. I'd like to collect a set of example positions like this to compare these moves. Once I have several such examples I may create a new post.
Black is pushing from behind. How should white respond?
The answer is the one-space jump at a. This move is faster than a nobi, but it's important to check that the weakness of the one-space jump is not exploitable:
Running technique Solution follows:
I've previously discussed how this is a nice shape. The 3-3 invasion at a is also a good option. My move was the one-space jump at b, but this is inferior by about 0.4 points.
Actually every white move here is a mistake.
should perhaps be at a. But I will especially remember that is aji keshi. Black will connect against the peep, which removes the possibility of a later atari by white stones at b and c. Instead, the kosumi-tsuke at d is a tesuji to make shape.-
hakuseki
- Dies with sente
- Posts: 100
- Joined: Thu Oct 29, 2020 6:18 am
- Rank: KGS 2 dan
- GD Posts: 0
- KGS: hakuseki
- Been thanked: 15 times
Ten more openings
I've played another ten openings. This time my average loss was about 12.88 points. This is 4.01 points worse than my result in my previous set of 10 games. I might attribute this apparent step backwards to such causes as:
Too much time spent reviewing instead of playing.
Focusing less on defense. In my previous set of games, I made few mistakes on defense, which led me to think I didn't need to focus on it as much. But after changing my focus, I am making more defense mistakes.
Frequently ignoring approaches to approach a different corner. This is not wrong but may lead to a more complex opening with more mistakes.
Playing while tired.
Too small sample size.
Anyway, I would like to try to replicate my previous result of 8.87 points lost, on average, in the opening. That's already a big improvement compared to a month or two ago, so I would be happy if I could consistently achieve it. I will try following the same checklist I used in that set, and play more games, perhaps 30.
Too much time spent reviewing instead of playing. Focusing less on defense. In my previous set of games, I made few mistakes on defense, which led me to think I didn't need to focus on it as much. But after changing my focus, I am making more defense mistakes. Frequently ignoring approaches to approach a different corner. This is not wrong but may lead to a more complex opening with more mistakes. Playing while tired. Too small sample size.Anyway, I would like to try to replicate my previous result of 8.87 points lost, on average, in the opening. That's already a big improvement compared to a month or two ago, so I would be happy if I could consistently achieve it. I will try following the same checklist I used in that set, and play more games, perhaps 30.
-
hakuseki
- Dies with sente
- Posts: 100
- Joined: Thu Oct 29, 2020 6:18 am
- Rank: KGS 2 dan
- GD Posts: 0
- KGS: hakuseki
- Been thanked: 15 times
Results from 64 more openings
I haven't posted recently, but my studies continue. I've played a new set of 64 openings, moves 1-46. Here are my results:
As mentioned, I was in a bit of a slump previously, but I seem to have recovered and maybe even improved a bit. Some observations:
I'm not spending as much time reviewing now. Reviewing is still important, but I think I should spend more time playing the game than reviewing it.
I am trying to play to my strengths rather than strictly following the AI style, especially in the very early moves. I think this leads to positions that I understand a bit better and results in me losing fewer points in the late opening.
I am paying attention to the "Adjusted Point Loss" which is calculated as my point loss minus 1/2 of my opponent's point loss. Essentially, I am taking credit for causing half of my opponent's mistakes. This lets me play a bit more adventurously when I feel so inspired. For now I have reported my point loss as well, for comparison to earlier results, but I will stop tracking this in the future.
I am still studying group tax-style play, but I have stopped forcing the traditional cross-hoshi setup. I do play cross-hoshi at every opportunity, but my opponent has the option to decline now.
I'm setting up 8 different types of games: me playing as black or white; either the full KataGo engine or just the policy model for my opponent; either stone scoring or area scoring for the rules. I am always reviewing with stone scoring even if the game was played under area scoring; these variations are just to induce a variety of play styles from my opponent.
Code: Select all
Point Loss Adjusted Point Loss
Games 01-08 16.7 13.1
Games 09-16 12.7 12.4
Games 17-32 12.8 11.2
Games 33-64 8.3 6.9
-
hakuseki
- Dies with sente
- Posts: 100
- Joined: Thu Oct 29, 2020 6:18 am
- Rank: KGS 2 dan
- GD Posts: 0
- KGS: hakuseki
- Been thanked: 15 times
Tsumego Practice
I've been practicing life and death using the BadukPop app. I upgraded to a Pro account so I can study problems with no time limit.
On September 1st of this year, I measured my time to complete a set of 5 problems at each difficulty level 1 through 5. There's also a level 6, but I did not attempt those problems as they felt very hard to me.
Since then, I have been drilling the level 4 problems every day. Today I achieved a time of 10.3 seconds per problem, measured over 15 problems. This is less than my original time for level 3 problems, so one might say that I have improved by one level, which was my goal.
To prove that I have advanced by one level, I will next need to achieve 25.3 seconds on level 5 problems and 58.5 seconds on level 6 problems. I will post later on how that goes.
On September 1st of this year, I measured my time to complete a set of 5 problems at each difficulty level 1 through 5. There's also a level 6, but I did not attempt those problems as they felt very hard to me.
Code: Select all
Level 1: 3.4 seconds per problem
Level 2: 6.5 seconds per problem
Level 3: 10.9 seconds per problem
Level 4: 25.3 seconds per problem
Level 5: 58.5 seconds per problem
To prove that I have advanced by one level, I will next need to achieve 25.3 seconds on level 5 problems and 58.5 seconds on level 6 problems. I will post later on how that goes.