bug-gnubg
[Top][All Lists]

Re: [Bug-gnubg] pubeval benchmark

 From: boomslang Subject: Re: [Bug-gnubg] pubeval benchmark Date: Fri, 3 Feb 2012 16:15:50 +0000 (GMT)

Hi Mark,

I do use the race/contact nets based on the type prior to the move, so i think i implemented that correct.

I wondered if gammons matter, because while reinforcement learning, moves are made.  For me, these moves are made based on maximizing winning probabilities; when you take gammons into account, I assume you make moves that maximize "prob(W) + 2 * prob(GW)", so different moves are made.

Anyway, it might be senseful to check my PUBEVAL implementation once again.

Ciao

Sent: Friday, 3 February 2012, 17:00
Subject: Re: [Bug-gnubg] pubeval benchmark

If you're just looking at probability of win the gammon node doesn't matter, though of course if you want to look at equity then you'll get value from it.

Using 80 hidden units and simple inputs I got a player to 63% wins against pubeval, and gnubg 0-ply (with more hidden units and extended inputs) wins around 71%.

So 67.5% sounds a bit high but not unbelievable. (On 20k matches the standard error on the % of win estimate is around +/- 0.35%, so 67.5% is significantly different from 63% or 71%.)

It's surprisingly tricky to implement pubeval correctly - I had a bunch of mistakes in my first attempt, and gnubg's implementation also had a bug until recently.

One subtle implementation point: when comparing potential moves you have to make sure that you use the race or contact weights based on the starting position, not on whether each potential move is contact or race. That's because pubeval's evaluation function is a separate linear regression for contact and race, and so the results of the two regressions aren't sensibly comparable (ie they don't represent probability of win).

That makes a smallish but noticeable different to average pubeval performance.

On Feb 3, 2012, at 10:32 AM, boomslang wrote:

i have a net that won 67.5% out of 20k matches.
It has 40 hidden units and relatively simple inputs (dummies for 1, 2 and more than 3 stones, and an integer for the excess of 3)

note: it doesnt have a notion of gammons yet.  Does this make it less comparable?

gr boomslang

Sent: Tuesday, 17 January 2012, 6:28
Subject: [Bug-gnubg] pubeval benchmark

How does gnubg perform against the pubeval benchmark in cubeless play?

I ask because I'm playing around with a backgammon network and have got one that wins 83% of games and +0.945ppg against pubeval (10k cubeless games). This is a single 80-hidden-node network with outputs for prob of win, prob of gammon win, and prob of gammon loss; and just the original Tesauro inputs. 0-ply.

But in the TD-Gammon scholarpedia article it says that TD-Gammon 2.1 in 1-ply mode wins only +0.596ppg against pubeval. (I think 1-ply here means the gnubg 0-ply.)

http://www.scholarpedia.org/article/Td-gammon

That seems really low compared to my result, since I'm pretty sure 2.1 had gammon outputs and also extra customized inputs.

So I'm wondering if I'm interpreting this correctly, or if I have an incorrectly-setup version of pubeval, or something like that.

_______________________________________________
Bug-gnubg mailing list
https://lists.gnu.org/mailman/listinfo/bug-gnubg