Potential of TensorRT+60block Network

[michito744]

I usually make a lot of hard requests, but I would like to report that the combination of the latest 60b and TensorRT versions is good for "searching long straight route". However, none of the contents of the configuration file for analysis matches the default_gtp.cfg.:-)

Note: I set komi to 7.0 in order to make W take it seriously(in Chinese rule).

Lee Sedol vs AlphaGoLee 5th game:

It is very good that small number of searches brings up this candidate.

B is right in deciding that it is impossible to move S8.

He is not happy with the progression above, and so W explore of moving S7 first. It's very good.

Since it is impossible for B to move S8 even with the above progression, KataGo finally come to the following conclusion. This is probably best. Which is the first time KataGo has reached this conclusion in so few searches.

Below is actual progress. In the old 40b network, it was computationally more expensive than this just to come to the right conclusion at this phase. AlphaGoLee was ridiculed by high-level human players because he could not make the right moves in this phase. I assume that this is due to both the fact that it had a margin in the actual game at 7.5 komi, and the fact that the AI is actually that much more advanced now from that time.

[michito744]

[kata1-b60c320-s4847033088-d2533443523]

A) It is excellent to be able to detect it at this stage.

B) Exploring 1 million times in the difficult phase of a very famous game. 1st assumption diagram: 2nd assumption diagram: Very excellent! And Mr. Lee was a true demon lord:-)