Life In 19x19

Bill Spight wrote: 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?

If a human plays just a few games against such unevenly skilled bots, he may find himself in a depressing sitation somewhere in the middle game and resign. Not realising that the bot is likely to trip and fall in a complicated fight later in the game.

But humans who play a lot against such a bot will pick out its weaknesses and exploit those. Using this they can inflate their rank quite a lot. It is not allowed on OGS because it has a destabilising effect on the rating system, but it has happened multiple times already. In some cases players were banned for this.

To avoid this problem, there are proposals to make all bot games on OGS unranked, like: https://forums.online-go.com/t/proposal ... ated/27543

gennan wrote: If a human plays just a few games against such unevenly skilled bots, he may find himself in a depressing sitation somewhere in the middle game and resign. Not realising that the bot is likely to trip and fall in a complicated fight later in the game.

I don't think it is that likely that a bot will mess up a game where the bot is far ahead (abeit still much more likely than a human player to do), and typically a human player (up to mid-dan level) will only resign when a large group has died.

gennan wrote:

Bill Spight wrote: 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?

If a human plays just a few games against such unevenly skilled bots, he may find himself in a depressing sitation somewhere in the middle game and resign. Not realising that the bot is likely to trip and fall in a complicated fight later in the game.

The fight doesn't necessarily have to be complicated. Once upon a time, there were MCTS bots rated around 3 kyu that frequently made DDK errors at the dame stage. But their human opponents often did not fill all the dame, not realizing the bots' weaknesses there.

gennan wrote:But humans who play a lot against such a bot will pick out its weaknesses and exploit those. Using this they can inflate their rank quite a lot. It is not allowed on OGS because it has a destabilising effect on the rating system, but it has happened multiple times already. In some cases players were banned for this.

To avoid this problem, there are proposals to make all bot games on OGS unranked, like: https://forums.online-go.com/t/proposal ... ated/27543

I want to know how much stronger I can become if I improve my reading. Here by reading I strictly mean the ability to visualize the board after playing an imaginary sequence. Usually after playing out around 5-6 moves I start to lose track of the stones that I was playing in my head or miss some key weakness. I am wondering if it would make a big difference if I could track the moves perfectly - that is if I could play it out on a real board or with an offline analysis tool. To test this, I am going to play OGS ranked games on two different OGS accounts. On one account (my main account feiqi__) I will not use any sort of analysis-aid, which is how I normally play anyway. But on a separate account (test account named speck_test), I will aggressively use the OGS analysis tool and try to play out future moves on a board whenever possible. On both accounts the opponents will be restricted to OGS bots (mostly doge_bot_2 and Spetral-1k). Let's see after 30 ranked games on each account if there is a difference.

The experiment started with 2 ranked games doge_bot_2 on my main account. So far I am 1-1 against doge_bot_2 but that winrate is unlikely to hold. In fact I probably will only have a 25% winrate against doge_bot_2, as I am rather weak against its style (probably the main advantage of doge_bot_2 over Spectral-1k is that it moves slower, which likely suggests that doge_bot_2 is set to do more Monte-Carlo playouts, which likely implies better reading).

Here is my first ranked win against doge_bot_2



I messed up in the opening, and somehow I always fall behind in the opening despite the very unusual attachments to 3-4's. OGS AI had white at 98+% fairly early and pretty much throughout the entire game. Eventually a miracle happened and the bot messed up

Surprisingly my win-rate against doge_bot_2 has drastically improved as I started to play ranked games. While I'm not quite at 50% against it yet, I'm winning a lot more than the 25% that I had initially projected, and OGS is giving me a rank at 2k based on mostly doge_bot_2 games.

Today I played my old "rival" Spectral-1k again, and I made one of those out-right blunders. It's a shame because I thought I actually played pretty well up until that point.

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X X . . |
$$ | . . . . . . . . . . . . . X . O X O . |
$$ | . . X , . X . . . , . . . . O X O O . |
$$ | . . . . . . . . . . . . . . X X X O . |
$$ | . . . . . . . . . . . . . . . O O O . |
$$ | . . . . . . . . . . . . . . . . X X . |
$$ | . . . . . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . . O O X |
$$ | . . . , . . . . . , . O . O . , O X . |
$$ | . . . . . . . . . . . . . . O . O X X |
$$ | . . . . . . . . . X . X . X O X X . . |
$$ | . . . . . . . . . . . . O O X X . X . |
$$ | . . . . . . . . . . . . . X O X O O . |
$$ | . . . . . . . . . . . O 1 O O X X O . |
$$ | . . . O . . . . O , O . . O X X O O . |
$$ | . . . . . . . X X . X X X O O O X O . |
$$ | . . . . . . . . . . . . . X X O X O . |
$$ | . . . . . . . . . . . . . X . X X O . |
$$ ---------------------------------------[/go]

I still remember my thought process at this point:
1. Wow my key stones are in atari, I better capture at either N6 or P1.
2. Okay I will capture at N6. But wait... if I capture at N6 wouldn't my M5 stones get cut.
3. Okay but so what? I can probably just sacrifice M5 stones anyway. After all I had just won a huge battle on the right side. Black is supposed to get at least a little bit of compensation for losing a big group...
4. Is there way to make sure that M5 remains connected? Oh I see if I play at M6 I can create a connect-and-die shape. How neat!
5. But wait, why do I even care if black cuts off M5, isn't this game already over anyway? Just play N6 to capture!
6. But if I could connect and save M5 so easily, then why not?
7. Okay I will play at M6.
8. OHHH OOPS WHY DID MY STONES DISAPPEAR

So I was debating with myself whether it is acceptable to give up on M5 stones completely, I started to forget the very first point- that my stones are in atari.

While I don't make these types of blunders too frequently, I do wonder if I get stronger these blunders will just disappear. But then I came to the realization that even when I was a high school student I could not find a way to eradicate those simple careless errors from those easy trigonometry/calculus exams (despite the fact that I was capable of solving more difficult problems in math contests). I should just accept that such mistakes will happen occasionally and move on.

Yes, blunders like this will happen in human play. Even pros are not immune to it.

There is something weirdly beautiful about this Quixotic quest to bring down a single in-human opponent. At least you know that your chosen giant is powered by sails...