[Computer-go] How to Research Brilliantly?

[David Ongaro]

Am 31.12.2010 um 08:37 schrieb Robert Jasiek:

> On 31.12.2010 03:16, Fuming Wang wrote:
>> This is certainly a good time to sit back and look at what got us here. The
>> following key ideas have been mentioned so far: UCB, MCTS, RAVE, Pattern and
>> Go knowledge during MC simulation.These ideas are all essential to a strong
>> MC based Go program.
> 
> So how did the theorists invent their ideas? Was it non-go research at first or did go players invent the relevant theories? How do we find more such good theories? Is that as hard as inventing mathematical descriptions of human-like strategy, for which one needs to study several years before creating something new looking brilliant, or are there more efficient ways for research of computer go theory?

There's surely no single cause for the recent improvements. (But I guess you expected no simple answer anyway...)

CGOS is an invaluable Infrastructure to provide statistics to do actual measurements. gogui helps to avoid lots of NIH and spares developer resources. GTP is developed further and gets wider adopted. Many small things, but combined they have quite an impact.

But most intriguing for me is the improved systematic approach to Computer Go in recent years. It's great to see more and more arguments backed up by statistics in this list. And to be honest, a lot of bullshitting went on in this list. And way too much time went into the development of expert systems, which where clear to be a dead end (even though they where the best we had for a long time). 

So if you're going to ask where do all these research ideas come from, the answer is simple: by *doing good old systematic research* at all. That is not to say there was no Computer Go research before 2006, but as we all now all these academic papers didn't provide a strong Go program. But this changed when Rémi published his MCTS paper in 2006 and a Go Program (CrazyStone) that could potentially outperform an expert system like gnugo.

Therefore I don't think the main achievement of Rémi is just bringing MCTS to attention to the Computer Go developer community. Much more important: he brought the academic systematic approach to the practical down-to-earth Computer Go day-to-day development.

Some say the discovery of MCTS had to come this late, because it needs computer power to be efficient. But I think this misses a great deal of the story. I'm sure todays MCTS programs can beat the hell out of expert system Go programs even on year 2000's hardware. Even though the time was ripe in 2006, I think with a more systematic approach in Go we could have get there much earlier. Therefore Rémis achievement cannot be overestimated.

If we're honest most of us didn't believe a MC based approach can work in Go (even in 2007 Darren Cook tried to argue the success of MoGo was due to UCT not to MC http://dcook.org/compgo/article_the_problem_with_random_playouts.html). Such non-intuitive results can only be achieved with systematic research. And if we look back in history, the greatest insights in all kinds of sciences are non-intuitive. And these are also the ones that lead to great progress.

If these improvements in knowledge-, tool-, protocol- and infrastructure-sharing continues, in addition to a systematic scientific approach, I'm really looking forward in what the future will hold for Computer Go.

I think it has just started.

David


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20110102/aff17aab/attachment.html>