Re: [Bug-gnubg] Questions on pruning net

[Joseph Heled]

Ian Shaw wrote:
> 1. When the pruning nets have finally selected a move, is that move
> re-evaluated using the full nets at the relevant play, to give the best
> possible evaluation of the equity?
I think most of you still miss the point of the prune nets. It just pre-filters 
the options in a move selection *inside a higher ply*. If the move is on it's 
top 10 list, the error is 0. If not, the error is the equity difference between 
the top move on the list and the top *0 ply* move.

> 2. How strong are the pruning nets in their own right? How do they
> compare with 0-ply?

Here is the benchmark results

contact

prune errors 49562 of 107485 avg 0.0350641438384

37299 errors of 107279
cube errors interesting 37299 of 107279
  me 0.0607724176278 eq 0.00221270350169


0p errors 32961 of 107485 avg 0.0104453134853
25395 errors of 107279
cube errors interesting 25395 of 107279
  me 0.00222095930382 eq 0.000191819048294


race:

prune errors 22471 of 96620 avg 0.00182693037711
10201 errors of 119067
cube errors interesting 8502 of 105573
  me 0.00172963250075 eq 0.000259362618432


0p errors 16847 of 96620 avg 0.000588811408958
7478 errors of 119067
cube errors interesting 5688 of 105573
  me 3.72035917218e-05 eq 3.79349081769e-06
cube errors non interesting 1790 of 13494
  me 0.000114145198118 eq 0.0

So, prune error rate per move is ~ three time higher than 0 ply.

"I don't understand those numbers", I hear you saying very quietly. So
In money play, prune give around  0.50960 ppg to 0ply.
In 7pt matches, 0ply wins approximately 80% of the time.


> 3. Might it be useful to use the pruning nets as a playing option lower
> than "expert", rather than introduce noise? I can think of two potential
> benefits. We might be able to get a really quick rollout - the current
> sub-expert settings are not useful at all, I think. Secondly, the
> pruning nets might have a more consistent playing style than the random
> effects currently used. This may give a "better" game than we currently
> have for lower level players.
I attached some 5pnt matches between 0ply and prune. You can make your own 
judgment ...

> -- Ian

pr.fibs.gz
Description: GNU Zip compressed data