Re: [Bug-gnubg] Wildly different evaluations by CRASHED network

[Joern Thyssen]

Hi,

On Sat, Jul 06, 2002 at 12:49:58PM +1200, Joseph Heled wrote
>  
> 
> I am fully aware of the deficiencies of the "crashed" net. Still, I
> think it is an improvement over the previous state. Not only does the
> crashed net understand "it's" positions better, the generic contact
> net is not burdened by understanding them.

What about adding a second or a series of crashed net(s) for handling
crashed positions with different number of chequers borne off?

> Is it possible to develop a better net(s)? Definitely. Can we do it? I
> am not sure.
> 
> The first hurdle, IMHO, is to develop a benchmark for comparing two
> nets to determine if one is better than the other. There is just one
> currency that counts in BG - making good decisions in move selection
> and cube handling. So, a good level 1(*) benchmark would be to take
> (say) 100K positions, rate the 5 best moves using a rollout, verify
> the correct cube decision with a cubefull rollout, and save those in a
> file. (the actual details and numbers, while tedious, can be worked
> out later). To do that we need a considerable amount of cycles - so -
> lets see how many mips can we come up with. I can offer around 700
> bogomips.

I can offer around 2000 bogomips -- but I'm not sure an intel-bogomip is
worth the same as a sun-bogomip :-) Meaning, sun probably does floating
point operations faster than intel.

When I return to work after the summer holidays I'll try to obtain some
CPU time from IBM.

Jørn

-- 
Joern Thyssen, PhD
Vendsysselgade 3, 3., DK-9000 Aalborg, Denmark
+45 9813 2791 (private) / +45 2077 2689 (mobile) / +45 9633 7036 (work)