Re: A very old bug reported on RGB

[Philippe Michel]

On Tue, Sep 27, 2022 at 04:21:52PM -0600, MK wrote:

> I'm sending you files that should help you even better. [...]

Indeed, the session sgf file you attached explains the issue (the 
gnubgrc files have nothing special).

The sgf file is the result of a session, but it is not really analyzed. 
It contains whatever evaluations the players used : 3 ply for chequer 
moves and either 3 ply or 0 ply + noise for cube decisions.

The latter isn't good, but what makes it worse is that there is only one 
record for cube decisions, even if it is displayed as two moves, and the 
evaluation recorded is the one from the last move, the take/pass 
decision. It clobbers the one from the doubling decision.

For instance, in game 1, player 1 (0 ply + noise) doubles at move 4. 
Player 0 (3 ply) takes and its evaluation is recorded : awful double and 
correct take.

A little later, at move 7, player 0 redoubles (right at 3 play) and 
player 1 takes ("evaluating" at 0 ply + noise the double as awful).

These cube actions (player 0 acts before player 1 whose evaluation gets 
recorded) are where the errors attributed to player 0 come from.

What you should do is analyze the session after it is over (with the 
"Analyse / Analyse match or session" menu) to have everything evaluated 
at a consistent level. If the analysis level is set to 3 ply it will be 
very fast: on average only half of the cube decisions and no chequer 
play will have to be reevaluated.

If you want to use another level (or be extra sure that the initial 
evaluations don't pollute the final analysis) you should use "Analyse / 
Clear analysis / Match or session" first.