[Pachi] Analyzing Pachi games

Petr Baudis pasky at ucw.cz
Wed Apr 20 01:55:34 CEST 2016


  Glad you like Pachi. :)

On Tue, Apr 19, 2016 at 07:32:21PM -0400, Evan Daniel wrote:
> So I'm just starting to explore the Pachi code base, and trying to
> develop an understanding of how the engine thinks. Any tips on what I
> should explore, or what I should do to analyze a game Pachi played?
> (Also, it's a neat engine, and I like the modularity, flexibility,
> available options, etc.)
> A very basic question: what sorts of win rates should I interpret as
> "ahead" "even" "behind" " farahead" etc.?
> For example, I played the attached game against Pachi. It was running
> on my slightly older desktop, at ~ 20k playouts/move. (Default timing
> settings? I just fired up GoGui.) It seemed like even fairly late into
> the game, Pachi didn't realize how much trouble it was in (winrate >
> 0.40).

  In general, in an even game the winrate should coast between say 0.51
and 0.53 (roughly) - above 0.50 as the total winrate is formed by
exploring favorable looking branches in the tree.  Below that already
signifies some trouble.

> I played the game without the pattern files (
> http://pachi.or.cz/pat/gogod-handikgspachi-iter/ ), but they don't
> seem to make much difference. (See attached analyse-sgf.pl output.)
> So, if I wanted to understand what went wrong, what would be a good
> starting point? (Opinions about what went wrong are also welcome, but
> I'm more curious about the methods you'd suggest using than the
> specific answer.)

  In general, the key debugging tool is the text board Pachi dumps
after every move. :)  The right part of the board denotes expected
territories and if Pachi is misreading some group, the expected
territory won't be what *you* expect.  Next step is looking at the
dumped game tree for sequences pertaining that group, but often
enough Pachi will try to avoid reading the group too much since
that tree branches aren't favorable.  So the key to the reading mistake
is then in the playouts, which means loading the position into Pachi
(e.g. using sgf2gtp) and running it with -d 7 or higher to see
individual playouts and look what happens.

  Kind regards,

				Petr Baudis
	If you have good ideas, good data and fast computers,
	you can do almost anything. -- Geoffrey Hinton

More information about the Pachi mailing list