[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] Confidence intervals from rollouts

From: nis
Subject: Re: [Bug-gnubg] Confidence intervals from rollouts
Date: Wed, 04 Sep 2002 11:31:05 +0200

--On 03  Sep 2002  19:37 -0700 David Montgomery <address@hidden> wrote:

to overestimate the standard error. Has anyone determined how much the
overestimate is?

It depends on the position, of course.  Let's assume
positions like the openings and responses.

Long ago I did some calculations based on JF rollouts,
which use stratified sampling.  I got a negative result
-- the data said the differences were larger than you
would expect from truly independent random samples,
rather than each rollout using stratified sampling.

Could you elaborate on which experiments you did, and which calculations you did on the results? Note that stratified sampling (at least in the form of perfect distribution on first plies) _should_ increase the standard error as calculated by gnu.

I'd have more confidence in my result if someone
repeated the experiment, however.

I will try to think of which experiments would make sense - something like doing a huge (stratified or not) rollout, then comparing the results to small stratified or unstratified rollouts.

Is it possible to do unstratified rollouts with gnu? And can anyone tell me what the "Rollout as initial position" does?

Ah, you just stipulate to cycling through the opening
ply or two.  JF's and my rollout code actually does more
than this, ensuring both perfectly distributed sampling
of the first two ply, and duplicate dice for subsequent ply
for every game.  That's what I mean by stratified sampling.

Just to make sure: "My rollout code" above means the one in gnubg, right? Could you elaboreate on what "ensuring duplicate dice for subsequent ply" means?

This should be slightly better than just distributing
the first ply or two, although to the degree there is
any benefit, the first two play probably captures most
of it.

I agree, especially when combined with the luck-based variance reduction.

Nis Jorgensen

reply via email to

[Prev in Thread] Current Thread [Next in Thread]