Endgame theory
elpanor opened this issue · comments
- In end game ..sente and gote is important
But Ai don recognize sente and gote.
Ex) after move_cnt 200, compare candidate 1 2 3
Candidate 1 a3 a4 b3 b4
Candidate 2 b3 b4 a3 a4
Candidate 3 .......
Then i choose candidate 2 and compare win_rate
(a3 a4 b3 b4 //// b3 b4 a3 a4)
If b3 b4 a3 a4 win_rate is higher than candidate1 sequence ..then candidate 2 was sente.
This methods may cause speed down.
- 2.11 has bug(??) for all shapes.
They recognize not only empty vertex but also placed stone for eyeshape falseeye selfatari and so on.
Is it intended?
- 2.11 has still don recognize some shapes(ex flowe five ) so i am checking for unusual types with making that types in tree search.
If it is difficult to check most types I think Aq needs to make special unusual cheking system.
- I don know it is possible.
Aq has errors in 1line and corner Nakade & vital.
I think separate nakade & vital point in mid position with 1line and corner position.
Corner position vital point is much different in mid position.
4.. ex) in mid position straight six is alive type. But in corner much more complex.
AQ is open source software, anyone can participate in development. If you feel that you can make AQ stronger, implement it, build it, measure the rating difference with the current version, and send a pull request.
Hi, elpanor. I'm sorry to have a question not related to this article.
Did you build AQ on windows? Which version of tensorflow did you use for building it on VS2015?
@trainewbie I built it with contrib/cmake in TensorFlow r0.12. I think later versions (r1.3 or r1.4) may also work.
@ymgaq Thanks for your kind answer 👍 :)
This issue is closed due to lack of recent activity.