I have returned to go recently and I am addicted to playing Spectral-1k on OGS. Although the name of the bot has "1k" in it, its rank on OGS is often 1-2d. Over the past few weeks, I have managed to play 87 games against it. My current score is 29 Wins - 58 Losses. My goal is to get to 50% win-rate or hit one thousand games. One thousand games is actually quite attainable because the bot moves very quickly (I generally play slow but even then these games usually finish in 30-45 minutes). This bot is can be very strong but has some rather hilarious blind spots. It seems quite likely that I will at least get to 30 wins before game 100.
Here is the most recent game I have played.
Some highlights :
This position probably highlights my biggest weakness. Katago analysis shows that the fight in the center is sufficiently settled, and therefore this is the opportunity to take the biggest territory move. For some reason I was scared of white cutting so I played an awfully slow hane with very little point value. Moves like this is the reason why I have switched to a more fighting-oriented style and why I try to avoid peaceful building games against this bot.
As a result I was on track for a close-loss until this happened.
The bot played to capture two stones at (a) but lost the entire groups in the center. Overall I got pretty lucky to pull this out.
A Thousand Games with Spectral?
Re: A Thousand Games with Spectral?
I had a very rough week with Spectral. By now I played over 100 games and my record is 36 Wins - 81 Losses (I only won 7 out of 32 games since my previous post). After a bad four game losing streak I think I went on tilt and had something like a 12 game losing streak. I am thinking about playing the next 50 games with emphasis on reading. Maybe it will be easier for me to improve my reading rather than positional judgement/evaluation.
Here is a game where I just completely collapsed due to poor reading.
Katago says the game was still roughly even until this losing move. I thought I read a sequence to capture black but it really didn't work. By the time that I realized that it didn't work it was already too late to pull back.
Here is a game where I just completely collapsed due to poor reading.
Katago says the game was still roughly even until this losing move. I thought I read a sequence to capture black but it really didn't work. By the time that I realized that it didn't work it was already too late to pull back.
Re: A Thousand Games with Spectral?
I have just improved my record to 40 Wins - 82 Losses; somehow I got a little bit of a winning streak going. I want to share a game that I played last week, which I thought was one of the more exciting games that I have played.
Basically I got to a rough start with some poor decisions. First I thought making a big wall here would worth something, but AI(either Leela Zero or Katago) disagrees and already gave white a 95% winrate.
But it got even worse : here for some weird reason I could put those stones in a net and create some thickness in the bottom right, but there was no net. Of course what happened after was a complete disaster
I managed to make a ko in the bottom right and pursued a desperate strategy to play out a broken ladder (in fact nonexistent) on the top right white stones. For some reason the bot was not willing to capture at (a), so this strategy does give me virtually an unlimited number of ko threats.
Finally black started to show some signs of life but AI says the game remains difficult for black.
The comeback became very promising at this point though
In the end the Spectral bot showed its key weakness in missing this huge atari (it seems like Spectral bot can't keep track of liberties for large groups). But even if Spectral did not make this silly mistake, just killing the white group in the center would have been enough.
This is one of the most improbable comebacks that I have completed.
Basically I got to a rough start with some poor decisions. First I thought making a big wall here would worth something, but AI(either Leela Zero or Katago) disagrees and already gave white a 95% winrate.
But it got even worse : here for some weird reason I could put those stones in a net and create some thickness in the bottom right, but there was no net. Of course what happened after was a complete disaster
I managed to make a ko in the bottom right and pursued a desperate strategy to play out a broken ladder (in fact nonexistent) on the top right white stones. For some reason the bot was not willing to capture at (a), so this strategy does give me virtually an unlimited number of ko threats.
Finally black started to show some signs of life but AI says the game remains difficult for black.
The comeback became very promising at this point though
In the end the Spectral bot showed its key weakness in missing this huge atari (it seems like Spectral bot can't keep track of liberties for large groups). But even if Spectral did not make this silly mistake, just killing the white group in the center would have been enough.
This is one of the most improbable comebacks that I have completed.
Re: A Thousand Games with Spectral?
Some random thoughts: Some time in July I decided to play go again. Honestly I am not exactly sure why I made that decision; I thought that I had stopped liking the game because it is so stressful to play. I started by making an OGS account and I played some of the bots on OGS. I quickly abandoned that account to start a new account where I would play Spectral-1k only. The first Spectral game was played on 7/25/20.
Some stats:
Up until now I have played 133 unranked games. My current record is 47 Wins - 86 Losses. On average I have been playing 5.54 games per day. I have won 21 of the most recent 50 games (42%). I have a conjecture that I have a much greater chance of winning if the bot plays a 4-4 in the opening (I may write a program to confirm this).
I am glad that recently I have been able to grind out close wins in relatively peaceful games. Here are two peaceful games (I won the first one but lost the second one).
The second was peaceful but not all that close (I thought it was close because I forgot who won the capturing race in the top left
).
Since my return to go I have not played a single human player. Perhaps one of these days I will find the very last game that I played against a human and review it here.
Some stats:
Up until now I have played 133 unranked games. My current record is 47 Wins - 86 Losses. On average I have been playing 5.54 games per day. I have won 21 of the most recent 50 games (42%). I have a conjecture that I have a much greater chance of winning if the bot plays a 4-4 in the opening (I may write a program to confirm this).
I am glad that recently I have been able to grind out close wins in relatively peaceful games. Here are two peaceful games (I won the first one but lost the second one).
The second was peaceful but not all that close (I thought it was close because I forgot who won the capturing race in the top left
). Since my return to go I have not played a single human player. Perhaps one of these days I will find the very last game that I played against a human and review it here.
Re: A Thousand Games with Spectral?
Overall I had a pretty good week. My current record against Spectral 1k is 55 Wins - 93 Losses. I have won 10 out of the last 17 games (and 24 out of the last 50 games). I don't know what is happening but when I first started this Spectral-1k series I couldn't even maintain a 15% win-rate. Somehow I have achieved statistically significant progress in a month. Either I am just getting better at abusing Spectral's weaknesses or I am genuinely getting stronger. As a perennial kyu-level player I am cautiously optimistic that spamming several hundred/several thousand games against a bot is the magical formula for success.
As I am getting very close to 50% win-rate based on the results of the last 50 games, I am not sure if it is better to keep playing against Spectral until I can dominate it (get to an 80% win-rate), or if it is better to try a stronger yet still beatable bot.
Anyway I still have a long way to go to get to an 80% win-rate. Here is a game that I just lost. It is one of the few games where I have lost due to poor endgame. I have always been reluctant to work on the endgame - it just seems very petty to find ways to squeeze a few points out of territorial borders. I have also heard that until you reach mid-dan level the best remedy to fix a bad endgame is to improve the middle game.
I was playing for moyo/influence but got into a difficult capturing race here
Complicated capturing race is often a good opportunity for me since I rely on liberty counting while Spectral only relies on the shape instincts of a neural net. There are plenty of situations where logic trumps instinct.
Eventually I got a ko in which black takes first, and because I didn't see any superb ko threats for white I was happy to play it.
Later I got a little passive (capturing races are exhausting on the human body so I did not want to fight anymore) so I just conceded the two stones to secure the connection in the middle - this was probably a mistake.
Eventually I got to this point, where the OGS AI gave me 99% winrate, but I managed to lose from here from poor endgame.
As I am getting very close to 50% win-rate based on the results of the last 50 games, I am not sure if it is better to keep playing against Spectral until I can dominate it (get to an 80% win-rate), or if it is better to try a stronger yet still beatable bot.
Anyway I still have a long way to go to get to an 80% win-rate. Here is a game that I just lost. It is one of the few games where I have lost due to poor endgame. I have always been reluctant to work on the endgame - it just seems very petty to find ways to squeeze a few points out of territorial borders. I have also heard that until you reach mid-dan level the best remedy to fix a bad endgame is to improve the middle game.
I was playing for moyo/influence but got into a difficult capturing race here
Complicated capturing race is often a good opportunity for me since I rely on liberty counting while Spectral only relies on the shape instincts of a neural net. There are plenty of situations where logic trumps instinct.
Eventually I got a ko in which black takes first, and because I didn't see any superb ko threats for white I was happy to play it.
Later I got a little passive (capturing races are exhausting on the human body so I did not want to fight anymore) so I just conceded the two stones to secure the connection in the middle - this was probably a mistake.
Eventually I got to this point, where the OGS AI gave me 99% winrate, but I managed to lose from here from poor endgame.
Re: A Thousand Games with Spectral?
Did I hit my ceiling in go? My current record against Spectral-1k is 83 Wins - 131 Losses. It looks like I have plateaued at around a 40% win-rate. I started to play against katrain-2k on OGS as well, and currently I am at 7 Wins - 5 Losses. If my record against these bots are to be believed, my playing level is around 1k to 2k. Considering that my record/win-rate has stabilized despite playing an insane number of games in 6 weeks, I can't help but wonder if I can still improve at this point.
-
HermanHiddema
- Gosei
- Posts: 2011
- Joined: Tue Apr 20, 2010 10:08 am
- Rank: Dutch 4D
- GD Posts: 645
- Universal go server handle: herminator
- Location: Groningen, NL
- Has thanked: 202 times
- Been thanked: 1086 times
Re: A Thousand Games with Spectral?
If you want to get stronger, you need to play stronger opponents. These bots are not playing dan level moves, so you're not learning them. It's much harder to surpass your opponents than it is to catch up to them, so it is to be expected you'd get to their level and then get stuck there.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: A Thousand Games with Spectral?
6 weeks is nothing at that level. Advancing 1 stone in 6 months would be better than most people do who get to your level.SpecK wrote:Considering that my record/win-rate has stabilized despite playing an insane number of games in 6 weeks, I can't help but wonder if I can still improve at this point.
How much are you reviewing your games? At your level I spent 1 hour per game on self review. Back then I was on my own. No bots. With help from a bot, maybe 30 min. per game would be good.
And yes, as Herman says, play against stronger players. The usual recommendation is to take 3 stones.
Congratulations on your record so far. Good luck!
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
Re: A Thousand Games with Spectral?
I used to review more diligently but recently I just take a quick 30 second look at the win-rate graph and some of spiky moves before hitting the rematch button.Bill Spight wrote:6 weeks is nothing at that level. Advancing 1 stone in 6 months would be better than most people do who get to your level.SpecK wrote:Considering that my record/win-rate has stabilized despite playing an insane number of games in 6 weeks, I can't help but wonder if I can still improve at this point.
How much are you reviewing your games? At your level I spent 1 hour per game on self review. Back then I was on my own. No bots. With help from a bot, maybe 30 min. per game would be good.
And yes, as Herman says, play against stronger players. The usual recommendation is to take 3 stones.
Congratulations on your record so far. Good luck!
-
Knotwilg
- Oza
- Posts: 2432
- Joined: Fri Jan 14, 2011 6:53 am
- Rank: KGS 2d OGS 1d Fox 4d
- GD Posts: 0
- KGS: Artevelde
- OGS: Knotwilg
- Online playing schedule: UTC 18:00 - 22:00
- Location: Ghent, Belgium
- Has thanked: 360 times
- Been thanked: 1021 times
- Contact:
Re: A Thousand Games with Spectral?
I don't have the magic key to improvement but if you play play play, it's very likely you will ingrain your current bad habits. Over-reviewing is the opposite trap but it may be good to have a regular deeper look at the major mistakes and discover possible patterns.SpecK wrote:I used to review more diligently but recently I just take a quick 30 second look at the win-rate graph and some of spiky moves before hitting the rematch button.Bill Spight wrote:6 weeks is nothing at that level. Advancing 1 stone in 6 months would be better than most people do who get to your level.SpecK wrote:Considering that my record/win-rate has stabilized despite playing an insane number of games in 6 weeks, I can't help but wonder if I can still improve at this point.
How much are you reviewing your games? At your level I spent 1 hour per game on self review. Back then I was on my own. No bots. With help from a bot, maybe 30 min. per game would be good.
And yes, as Herman says, play against stronger players. The usual recommendation is to take 3 stones.
Congratulations on your record so far. Good luck!
Re: A Thousand Games with Spectral?
I am adopting the principles laid out in this reddit post. https://www.reddit.com/r/baduk/comments ... advice_to/
The highlight (for me at least) of which is the following
The highlight (for me at least) of which is the following
Interestingly enough when I first started go I was the complete opposite of "play play play". I think I got to around 10 kyu by watching a lot of video lectures on strategy (back in the day there wasn't much go content available online so I actually just watched this one lecture series repeatedly) while playing well under 100 games.Playing is more important than doing literally anything else. If you want to dedicate a few hours to "learning", 9 out of 10 times I'd suggest playing a couple of games. It will be much more beneficial than learning a new joseki or a new sequence.
Learn them post-factum. When you encounter one or fail at one. Between 30k and 3k there are like 2-3 different joseki per corner stone (4-4, 3-4, etc) and you won't see anything else 90% of the time.
There is a major difference between playing 3 games a week and 7 games a week.
Many people would say that learning via "trial and error" is more time consuming, but I'll argue that. Polluting your mind with the concepts or sequences you can't grasp will cost you much more time than not doing so.
Re: A Thousand Games with Spectral?
This game felt like a complete beatdown from start to finish. I am not sure why Spectral-1k seemed to have played much weaker than usual. Typically even in the games that I had won I got into some very difficult positions.
Overall records
Spectral-1k 89 Wins- 139 Losses
Katrain-2k 8 Wins - 5 Losses
Katrain-2d 2 Wins - 3 Losses.
The katrain bots have been removed from OGS so it will be some time before I get to play them again (I understand that katrain is available for download but I don't know the exact settings to replicate the OGS version).
Overall records
Spectral-1k 89 Wins- 139 Losses
Katrain-2k 8 Wins - 5 Losses
Katrain-2d 2 Wins - 3 Losses.
The katrain bots have been removed from OGS so it will be some time before I get to play them again (I understand that katrain is available for download but I don't know the exact settings to replicate the OGS version).
-
gennan
- Lives in gote
- Posts: 497
- Joined: Fri Sep 22, 2017 2:08 am
- Rank: EGF 3d
- GD Posts: 0
- Universal go server handle: gennan
- Location: Netherlands
- Has thanked: 273 times
- Been thanked: 147 times
Re: A Thousand Games with Spectral?
Recently I reviewed a game of a human vs KaTrain "calibrated 1d", but that bot didn't play like a 1d human at all. Its positional intuition seems to be high dan level, but it makes tactical blunders more expected in human DDK games (for example this move).
From that game and the game you just showed (and your own observation in your 1st post "This bot can be very strong but has some rather hilarious blind spots"), I dear to say that the KaTrain bots around 1d seem to be quite broken and not playing similar to human players of that level at all.
So I admire the attempt of the programmer of the KaTrain tool to create these weakened versions of KataGo, but I would not recommend using those for training.
Why don't you play against humans?
From that game and the game you just showed (and your own observation in your 1st post "This bot can be very strong but has some rather hilarious blind spots"), I dear to say that the KaTrain bots around 1d seem to be quite broken and not playing similar to human players of that level at all.
So I admire the attempt of the programmer of the KaTrain tool to create these weakened versions of KataGo, but I would not recommend using those for training.
Why don't you play against humans?
Re: A Thousand Games with Spectral?
A computer which plays nothing like a human 1-d player may still be 50-50 against human 1-d players.gennan wrote:Recently I reviewed a game of a human vs KaTrain "calibrated 1d", but that bot didn't play like a 1d human at all. Its positional intuition seems to be high dan level, but it makes tactical blunders more expected in human DDK games (for example this move).
From that game and the game you just showed (and your own observation in your 1st post "This bot can be very strong but has some rather hilarious blind spots"), I dear to say that the KaTrain bots around 1d seem to be quite broken and not playing similar to human players of that level at all.
So I admire the attempt of the programmer of the KaTrain tool to create these weakened versions of KataGo, but I would not recommend using those for training.
Why don't you play against humans?
I have been playing Spectral-1k, which has tactical weaknesses (and probably even more serious ones than those of Katrain because of ladders). However Spectral-1k is remarkably consistent at holding a rank around 1k-1d on OGS, so the superior positional intuition makes up for the tactical blind spots. I have always been wondering if Spectral-1k and the Katrain bots are over-ranked, but I have seen the Spectral bot defeat some solidly dan-level players that I know. I do think the tactical weaknesses allow a much weaker human opponent to win at least 5-10% of the games.
I am not playing against human players because I want to test the hypothesis that it is possible to improve from playing a large number of games against computer opponents.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: A Thousand Games with Spectral?
This kind of combination of strategical strength and tactical weakness is something that I noticed with MCTS bots well before the advent of superhuman bots. Weaker bots of this ilk can blunder in the endgame, so my question about their ratings is this: Do their human opponents resign?SpecK wrote:I have been playing Spectral-1k, which has tactical weaknesses (and probably even more serious ones than those of Katrain because of ladders). However Spectral-1k is remarkably consistent at holding a rank around 1k-1d on OGS, so the superior positional intuition makes up for the tactical blind spots. I have always been wondering if Spectral-1k and the Katrain bots are over-ranked, but I have seen the Spectral bot defeat some solidly dan-level players that I know. I do think the tactical weaknesses allow a much weaker human opponent to win at least 5-10% of the games.
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.