question about opening book extension testing

[Uri Blass]

I wonder if people tried the following idea and if it is productive.

The idea is that a program play only with a small book that it trusts but has access to a big book in order to do opening book extensions.

If the program is out of the small book at move 10 it simply build a tree of all the games that people or programs played from the relevant position.

The tree may be limited to a smaller tree by conditions like
1)not to include lines of more than 20 plies unless there are at least 2 games in the relevant line
2)not to include lines when a move lead to clearly bad result based on statistics of significant number of games because these lines are probably not good.

Now the program simply extend every move in the tree by 1 ply(or by 1/2 ply)

The question is simply if the program with book extensions can get significantly better result relative to the program without book extensions based on games.

It is possible to test it also in CEGT or CCRL conditions because I assume that the games start from theory position that people played in the past
so one program is going to play with no book and no book move extensions and another entry may be the same program that does book move extensions.

I assume that building the relevant tree in positions when there are maybe few hundreds of games from the relevant positions should not take too much time.

[Uri Blass]

Note that I am not sure if condition 2 is good because
It is possible that a bad line based on statistics is simply a trap and it is better that the computer extend it in order not to fall into the trap.

[jdart]

It is worth a try.

As you note game results are not always reliable though because many games are lost for reasons that are not related to the opening (loss on time, blunder in the later middlegame or endgame, etc.).

--Jon