Poker-AI.org

Dear all,

I have been experimenting with tree solving for a while now and concluded that each of the sampling schemes has its advantages and disadvantages. In my mind roughly this is the case:
Vanilla - precise, but too slow and too big;
CS - accurate, but slow;
ES - good and reasonably fast;
AS - less accurate, but faster than ES.

But then I was thinking why can't we run start running the solving with Vanilla for 1HR then take the CU/AVG strats and use continue working on them with CS and afterwards pipe the results to ES/AS.

I have tried mixing two schemes, namely CS and ES but ES always outperforms the combo. ES/AS would be maybe even more attractive. Did anyone thought/ tried this and could share his thoughts?

_________________
Let's drop conventional languages and talk C++ finally.

Each algorithm has an associated convergence rate. Two cases:

[1] convergence rate(algorithm A) == convergence rate(algorithm B): you can happily mix them up, but you will not get any advantage from this
[2] convergence rate(algorithm A) != convergence rate(alrorithm B): either A outperforms B or B outperforms A. so there is no sense in mixing them up instaed of sticking with the outperforming algorithm

I can argue for a combination of AS -> Vanilla, because Vanilla does not have any variance associated