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

Re: [computer-go] Hardware-Instruction.



Don Dailey wrote:
Yes, I know computer chess has advanced. I gave up lazy eval before I
left computer chess because it wasn't safe unless I used big margins,
at which point it didn't help much at all.
I happen to disagree with this, but let's not get into that too much as it's not very interesting :)

But why MTD no longer used? Is it because MTD is so hard on the hash
tables and memory latency is a bigger issue? MTD is much more
dependant on the fact that the search tree remains mostly stored in
memory via the hash tables so I can believe this is possible.
I assume it's because the difference with PVS+narrow aspiration windows is not very large, many people use pruning or extensions that are alpha/beta dependant, and probing memory keeps getting more expensive.

When you say, "usually not near the leafs", what do you mean?  Are you
using R=2 near leaf or something else entirely or nothing at all?
Common are:

R = 3 at depth-left >= 6 ply, R = 2 below, if the program does not have a tactical quiescent search
R = 3 everywhere, if the program has a tactical quiescent search

But there were a number of  big problems we had during this tournament
which I think your comments are based on:

   1. Extremely poor internet connectivity.
That issue has only been fixed this year (2004), and it's still not very great - an infrastructure problem, partly.

2. Bugs. The user interface was incredibly buggy and really hurt
our program.
You were a n00b, in other words ;-)

      The tournament  director was inexperienced as he  made 2 errors,
That particular issue still isn't solved yet, it seems.

Anyway, you wrote a good introduction to Cilk, but I couldn't help to notice you didn't answer Vincent's inquiries about performance.

Please do so (and some supporting logs or whatever would be really great). The amount of verifyable parallel performance data in chess is close to zero and if you can provide some that would be quite interesting.

I'm tempted to see how quickly I can convert my current program to Cilk, and what the performance result of that would be.

--
GCP
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/