Re: [Bug-gnubg] Rollout problems

[Jim Segrave]

On Wed 04 Dec 2002 (19:50 +0100), Magnar Naustdalslid wrote:
> Thanks for nice new rollout features. I have encountered a couple of 
> problems:
> 
> Problem 1:
> Rollouts with separate evaluation for deeper plies are too slow. For 
> example, try the following: First, roll out the initial position untruncated, 
> expert level, 0-ply, 10 trials. This takes a few seconds. Second, roll out 
> the initial position at reduced 2-ply, medium search space, truncated at 
> one ply, 10 trials. This also takes a few seconds. Third, try to *combine* 
> these settings by evaluating the first move at reduced 2-ply, medium 
> search space and the rest of the moves at 0-ply expert. Strangely, this is 
> very slow. 

I just ran some tests with a random saved game file. I set the rng
seed to 0 for each trial (and each was done starting gnubg up again,
so no cached analsis would be available.

First, 4 untrucated 2 ply reduced rollouts from the initial roll 

3m 51 sec = 57 seconds/game

Switching to 0 ply evaluation after the first move:

1m 41 sec = 26 seconds/game

Switching to 0 ply evaluation after the fifth move:

1m 53 sec = 28 sec/game

Everything at 0 ply

5 sec = 1.25 sec/game


> Problem 2:
> GNUbg crashes when I try to run a rollout with reduced 2-ply, small 
> search space.

Yes - there was a bug. A fix was committed Tues 3 Dec.

-- 
Jim Segrave           address@hidden