[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] Moyoman, a Go playing program
On Jul 7, 2004, at 17:57, chrilly wrote:
1. It will take about 100 person years, or 1 million
executable lines of code to get to amateur 1 Dan.
I think 4-6 man years should be sufficient for this goal. 1 full-time
programmer, one part-time go-expert and one part-time tester and one - part
time - project manag. The goal should be reachable in 2-3 years.
In the computer-chess-project Hydra I have now these ressources.I think this
is the perfect project size. In a bigger software project the productivity
per member declines rapidly due to the communication/coordination overhead.
I concur. The average computer go player improves by something like one or two stones per year. For example, not too long ago, GNU Go was playing around 15k. Current version now plays 6-7k on NNGS.
Besides, "bigger is better" is a pretty foolish guideline for estimating the complexity of a given task. I actually expect Go programs to get much smaller (and smarter) in the coming years.
Also: Need to be reminded that progress on using neural networks to play board games has been tremendous in the past decade? It's only a matter of time before we stop coding everything by hand; instead, we'll feed up the program with data and it will learn on its own. That of course wouldn't require even nearly as much as 2% of that 1M lines of code.
Unfortunately I have found yet no sponsor who wants to invest this 4-6 man
years. Or to put it the other way round. Fortunately I have found a potent
sponsor for building the ultimate chess machine.
Strange...
Have you looked in Asia at the very least?
I have worked before in 1 million lines of code projects (big
telephone-system for Siemens). It is a nightmare. This amount can not be
reliable handled any more. Everyone hesitates to make changes to working
code (even if the code is bad and slow), because one can not overlook the
consequences of changes.
There is also the fact that these systems were mostly designed and implemented before I was even born. Back then, most designers probably didn't even hear about modular design, object-oriented programming, or perhaps even something as simple as code reuse. Every big old system is a mess because of that.
I also think this is the complete wrong way. I see no reason, why a
Go-Programm should be longer than 50k Lines. This is already quite big to
manage. If one needs more code, one simply has not identified the right
concepts. Adding to much knowledge results in a lot of white noise which
dominates in many positions the relevant features. One has also the problem
to make this knowledge consistent. In chess there is the law of diminishing
returns. Some basic (and usually also simple) features like e.g. mobility
contribute most to the playing strength.
Best Regards
Chrilly Donninger
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/