Passing Through (ireland.delete@this.web.ie) on August 3, 2018 4:07 am wrote:

> I've got a question that I've been trying to phrase about all of this too. Like back in 2003, one
> had people coding for chess artificial intelligence projects, and they were trying to use Intel's
> C++ compiler fifteen years ago, and not getting very far with it, trying to compile for use on
> ordinary single core x86 CPU's back then. I guess, one could have termed that very small sub-set,
> or niche community of developers back then who were trying to work with the Intel tools provided,
> as the early pioneers in artificial intelligence programming on the desktop products.
>
> But the question that I'd have, is that if those developers had Xeon Phi, or nVidia CUDA back then, would they
> have abandoned trying to work with CPU's entirely, and opted instead for a faster way to achieve their goals.
> Would they have abandoned trying to build chess AI on x86 processors entirely? Maybe, not because of the hardware,
> but because the compilers for the x86 processors at the time, weren't making it easy for them?

No, not at all. At the time general purpose processors were the fastest for regular heuristic alpha beta searchers. And they still are. And this is, today, still the best approach producing the strongest programs in fair conditions. Whether you use Intel C and it's auto-vectorization or not is largely irrelevant.

Most of the people working in this field hammer with the known tools. There is very few true research, and the computer chess field is mostly plagiarism.

Until the Google paper came out in 2017, people didn't even believe you could build a competitive chess program with DCNN + MCTS. And that's despite both approaches being proven in other fields. I own a book written in 2018 that still claims it! (No wonder, given the experts they cite...)

Now, in order for people to try this research, it helps if the tooling and hardware is widely available. Training Alpha Zero took 5000 special purpose ASICs. Still outside the real of possibility of all but a few institutions. That's why Google was the one to publish it first.

So, I'd say there is no clear cut decision based on hardware. But as a result of hardware being widely available and easily programmable, more research gets done. And sometimes research produces results.

One reason why so few research uses FPGAs, even though they were a good fit to this field: users don't have them, and the programming tools are flaming garbage.

> And turn that question on it's head to today. What if any reason would there be today for a member
> of the same kind of community of developers in artificial intelligence for chess games, to choose
> to develop their projects on top of generic x86 processor technology at all?

If you have a problem that's not amendable to DCNN's, why not?

DCNN are popular because they are, all things considered, easy, and often give good results. The workings of reinforcement learning with them are well documented and there is a lot of good software for it. You can get published when doing research with them. What's there not to like, eh?

But to think that this is now suddenly the only tool in the arsenal seems naive. For sure, the other tools are a bit on the back-burner due to, well, fashion.

> And there are other proprietary tools, that
> are intended to foster the development quickly of talent working around proprietary nVidia hardware.
> And while, I'm all for generic tools and means of accomplishing tasks in computers,... judging between
> the 'community' that is suggested by the Intel link overhead, or being a part of the community as is
> suggested by the nVidia link instead,... if you were a young kid now today in a post-graduate position,
> looking to actually finish writing your thesis, which one would you choose?

Fuck proprietary tooling.

--
GCP