[Computer-go] 7.0 Komi and weird deep search result
3-Hirn-Verlag at gmx.de
Wed Apr 6 06:55:51 PDT 2011
Don Dailey wrote:
> I would like to ask you to reconsider your feeling on this, because I
> they are misguided. Let me explain why I feel that way and you can take
> for what it's worth.
> Many months ago there was a discussion on computer chess and people were
> divided on the issue if most of amazing progress in computer chess was due
> to software or hardware. Most had always assumed it was hardware and that
> is the standard sound bite cut and pasted into articles that talk about
> progress of computer chess. However, I was swayed convincingly to the
> software side of the issue and I think most real experts on computer chess
> agree now that it's roughly even - the software progress in computer chess
> has been absolutely astounding when measure over decades and it has kept
> pace and perhaps even outstripped hardware.
> But here is the thing ... the whole process was long and gradual and it
> never involved any major breakthroughs. Yes, there were things that
> out and some refer to them as breakthroughs, but none of them were even
> fraction of what we got from Monte Carlo methods for Go. Go was really
> ripe for this one because it was long overdue anyway and global search was
> the only way forward and had been dismissed for so long.
> At each step along the way, it was almost impossible to imagine making a
> program play much stronger as it seemed like we were doing everything
> reasonably possible.
> In 2004 a program called "Fruit" came out and it astounded the computer
> chess community with it's incredible strength. The strange thing about
> this is that there are NOW chess programs that are hundreds of ELO
> than Fruit, running on the SAME hardware. It's not hardware, it's real
> software advancements, good ideas and good engineering. And there have
> been no "major breakthoughs" since Fruit to produce these programs. Even
> Fruit itself did not have any new ideas of major significance, it just did
> lot of things a little bit better. (Probably the biggest factor in Fruit
> was something called LMR which was worth something like 50 ELO, but even
> that was just an old idea all polished up with a better formulation.
> many think the major thing that made it strong was simplicity and bug free
> The lesson from this, is that if you are standing in a dense forest, you
> can only see a few trees around you. YOUR feeling that we have "almost
> reached our limit" and that we have exhausted all the good ideas in MCTS
> need something new is certainly shortsighted and I mean no disrespect. I
> have fallen for this kind of thinking a few times myself.
> What you are very likely going to see is gradual steady progress and
> constant refinements of ideas and incremental improvements which will add
> to an overwhelming amount of power in a relatively short time. There
> not be any major new ideas but probably a lot of little ideas, each
> contributing a small amount.
> Here is an example of how this works, using a principle every top chess
> program author understands very well:
> Imagine that you can make your program strong enough in a months time to
> 51% of the games against the previous version. If you can do that, you
> will have added something like 5-10 ELO to the strength of your program.
> 5-10 ELO is a small amount of improvement and it is so difficult to
> that it requires thousands of games to verify. So we are not talking
> about a herculean task. And yet, if you do that for a year you will
> have added 1 dan of strength to your go program! 1 Dan in 1 year is a
> But I don't think the engineers in Go programming think like that yet.
> (Maybe some do, I applaud you if you so.) They seem to get frustrated
> they cannot find an immediate big gain and speak of not making progress
> without some major breakthrough because they cannot imagine their program
> playing much better without one. The way you make a strong program is
> 1-10 ELO at a time over months and years. I think it is extremely
> that anybody will find a single 1 dan idea that is relatively painless to
> code up. However there are probably many good ideas that will add up to
> 1 dan with a lot of engineering and tuning and hard work.
> I admit that a big hurdle with GO is the ability to test 10,000 games (for
> instance) in a reasonable amount of time. It would probably tax the
> patience of most GO authors to try to measure small improvements and thus
> the mentality to only look for big things and dream of major
> But I seriously doubt anything that is quickly measured is likely to be
> found and I'm convinced that even the best ideas need a lot of refinements
> and engineering - the stuff that takes so long and requires so many games.
> So my predictions is that without anything really major (but no doubt some
> new small ideas) we are going to see your KGS 5 and 6 dan and much higher
> 5 to 10 years. I'm assuming that the good engineers in computer go will
> improve on their testing methodology (and I think multi-core newer
> hardware will make this much easier to do.)
Many small amens give one big AMEN.
Empfehlen Sie GMX DSL Ihren Freunden und Bekannten und wir
belohnen Sie mit bis zu 50,- Euro! https://freundschaftswerbung.gmx.de
More information about the Computer-go