Life In 19x19
http://lifein19x19.com/

Training plan
http://lifein19x19.com/viewtopic.php?f=48&t=16512
Page 3 of 3

Author:  jlt [ Tue Mar 26, 2019 3:40 am ]
Post subject:  Re: Training plan

I mean, 5 days ago your level was about 15k, and now even LeelaZero can't find any bad moves. That's amazing progress.

Author:  yakcyll [ Tue Mar 26, 2019 4:46 am ]
Post subject:  Re: Training plan

jlt wrote:
I mean, 5 days ago your level was about 15k, and now even LeelaZero can't find any bad moves. That's amazing progress.

Goes to show how much the right way of learning and the best teachers matter from the very beginning. Still on the edge about whether they are right and best in his case, but it's hard to argue they are not effective.

Author:  kaiq1a1 [ Tue Mar 26, 2019 4:56 am ]
Post subject:  Re: Training plan

yakcyll wrote:
jlt wrote:
I mean, 5 days ago your level was about 15k, and now even LeelaZero can't find any bad moves. That's amazing progress.

Goes to show how much the right way of learning and the best teachers matter from the very beginning. Still on the edge about whether they are right and best in his case, but it's hard to argue they are not effective.


What do you mean ''right way of learning and the best teachers matter from the very beginning''?? and I am finding his progress abit too quickly

Author:  Darknessmoves [ Tue Mar 26, 2019 7:40 am ]
Post subject:  Re: Training plan

yakcyll wrote:
jlt wrote:
I mean, 5 days ago your level was about 15k, and now even LeelaZero can't find any bad moves. That's amazing progress.

Goes to show how much the right way of learning and the best teachers matter from the very beginning. Still on the edge about whether they are right and best in his case, but it's hard to argue they are not effective.


Um, By teachers what do you mean???

Author:  jlt [ Tue Mar 26, 2019 1:46 pm ]
Post subject:  Re: Training plan

That LeelaZero teaches you efficiently?

Author:  Darknessmoves [ Tue Mar 26, 2019 10:30 pm ]
Post subject:  Re: Training plan

I've completed graded go problems volume 4 Got 74 wrong.

Did 1 game against leela zero. I am hanging on longer than I did before because my winrates is staying up longer.

Replayed 10 alphago games.

Did 2 games against humans lost the first one against a 6kyu won the second one against a 10kyu.

Nearly done with lee changho tesuji problems.

Went through direction of play but I didn't really understand it.

Here is the game I won against a 10kyu I was black.





Attachments:
darkness2.sgf [1.25 KiB]
Downloaded 517 times

Author:  jlt [ Wed Mar 27, 2019 1:00 am ]
Post subject:  Re: Training plan

Congratulations, the players almost look human. The opening lacks variety, though.

Author:  Darknessmoves [ Mon Apr 08, 2019 6:40 pm ]
Post subject:  Re: Training plan

New training plan now.

Graded go problems for beginners volume 4 redo.

Since I completed all alphago games excluding the self-play. I am going to replay 10 shusaku games.

Played against leela zero 13 times in the time I been gone.

Did 1 game against humans.

Completed lee changho tesuji problems volume 1.

Went through attack and defense again.

This is the game I played. I lost against 9kyu. I was black



Attachments:
dark17.sgf [2.43 KiB]
Downloaded 528 times

Author:  Uberdude [ Mon Apr 08, 2019 10:26 pm ]
Post subject:  Re: Training plan

What would you play if the position at move 36 was in the tesuji problem book?

Author:  jlt [ Tue Apr 09, 2019 12:02 am ]
Post subject:  Re: Training plan

Now the opening looks normal but the endgame is quite strange.

Author:  kaiq1a1 [ Wed Apr 10, 2019 5:28 am ]
Post subject:  Re: Training plan

jlt wrote:
Now the opening looks normal but the endgame is quite strange.


''Now the opening looks normal but the endgame is quite strange''. ???? your phrasing is strange?

Author:  jlt [ Wed Apr 10, 2019 5:55 am ]
Post subject:  Re: Training plan

Let me be more explicit then. In most of the games before the last one (or perhaps even all games), it is apparent that both players are bots, or at least humans selecting moves among bot choices. Here are symptoms upon which I made this judgment:

  • Clean josekis that are popular among recent bots, but which are not commonly used among DDKs or low SDKs.
  • Absence of really bad moves. All moves are one of LeelaZero's options.
  • Absence of typical DDK or low SDK mistakes, like bad shapes or lack of liberties.
  • Tenuki more frequent than typically seen in human games.
  • Early resignation, at a stage of the game when the winrate estimated by LeelaZero is low, despite being behind by 10-15 points or less.

The last game that was posted looked more human: the pincer josekis are more popular among humans than among bots. There are more shape mistakes, like the one pointed out by Uberdude. However there are strange moves at the end like 224 and onwards (why try to capture while connecting to the living group below would suffice?), 259 and 275 (adding a stone to a dead group?), 302 (fill an eye?), 319 and 323 (WTF?), moves 329--339 (WTF?). You only see these strange moves when bots are in desperate mode, or 100% confident they will win. I guess the resignation threshold was lowered to avoid early resignations (see above), however this results in non-realistic endgames.

Author:  kaiq1a1 [ Thu Apr 11, 2019 5:51 am ]
Post subject:  Re: Training plan

jlt wrote:
Let me be more explicit then. In most of the games before the last one (or perhaps even all games), it is apparent that both players are bots, or at least humans selecting moves among bot choices. Here are symptoms upon which I made this judgment:

  • Clean josekis that are popular among recent bots, but which are not commonly used among DDKs or low SDKs.
  • Absence of really bad moves. All moves are one of LeelaZero's options.
  • Absence of typical DDK or low SDK mistakes, like bad shapes or lack of liberties.
  • Tenuki more frequent than typically seen in human games.
  • Early resignation, at a stage of the game when the winrate estimated by LeelaZero is low, despite being behind by 10-15 points or less.

The last game that was posted looked more human: the pincer josekis are more popular among humans than among bots. There are more shape mistakes, like the one pointed out by Uberdude. However there are strange moves at the end like 224 and onwards (why try to capture while connecting to the living group below would suffice?), 259 and 275 (adding a stone to a dead group?), 302 (fill an eye?), 319 and 323 (WTF?), moves 329--339 (WTF?). You only see these strange moves when bots are in desperate mode, or 100% confident they will win. I guess the resignation threshold was lowered to avoid early resignations (see above), however this results in non-realistic endgames.


''Clean josekis that are popular among recent bots, but which are not commonly used among DDKs or low SDKs'' Well. it's possible to adapt those josekis if you study the kifu of the bots that play it.

''Tenuki more frequent than typically seen in human games.'' I don't feel like you can use this as proof, tenuki can depend on things.

''Early resignation, at a stage of the game when the winrate estimated by LeelaZero is low, despite being behind by 10-15 points or less.[/list]'' I've seen many ppl resign early. However I agree with you that this person is most likely selecting bot moves merely because of ''leela zero selected moves'' part I went through the games with leela zero just to confirm. and all of them are leela zero selected moves. it's next to impossible for a 16 kyu to do that many good moves without messing up.

Page 3 of 3 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/