It's a well-written inspirational/PR piece. Such things do have value. It doesn't claim to be presenting any novel facts, though your link looks like a good place to start looking for those.
One big question that interests me is:
how many possible positions does the game of Go have?
Answering that requires computing the 361st power of a
very sparse 363 billion by 363 billion matrix,
a big computation indeed (see http://www.cwi.nl/~tromp/go/legal.html for details).
Right now, I'm computing the number of 18x18 positions
(standard size is 19x19) using a couple of servers with
512GB memory and a few TB of disk-space, and that's taking almost a year.
I'd love to have access to a super-computer, but they tend to frown upon such frivolous computations...
But does that _require_ big computation, or aren't we smart enough to do this the easy way?
I know simple induction will not work for this problem, but to me, this feels like the kind of problem that might be simple enough that some mathematical breakthrough could make it a lot simpler.
Sometimes with this sort of thing you can exploit symmetry in the game board to reduce the amount of work you need to do. I don't know enough about Go (or anything at all really) to say if that's possible or not in this situation, but it is probably worth looking into.
I agree though, this does seem like the type of problem that should have a relatively simple solution, or at least a reasonable approximation.
We have excellent approximations; for 19x19 the number is known to be approximately 2.081681994 * 10^170. Getting the exact number is the big challenge.
You can see in the paper that we spent a lot of effort on reducing the complexity of the counting problem. We welcome suggestions for further improvement, but suspect our current algorithm (counting paths in the border state graph) is already close to optimal.
Many scientific disciplines require supercomputers with fast interconnects for their complex simulations. Genomics is not presently one of them, we just need lots of compute time on fairly normal servers.
The HPC does get used of course, but not for purely technical reasons. Genomics workloads should move to the cloud and be run at off-peak times to take advantage of cheaper pricing, but there are significant barriers to this as researchers can often get started on the HPC for free whereas cloud services require a purchasing card. Their cluster will also have some support and someone who can install software dependencies into the stack.
(As a result most genomics software is really, really hard to get running on anything other than the author's institutions HPC and its unique set of installed packages.)
Another barrier is just knowledge (it's hard to keep those people around in the current market) and university projects often have strange incentices due to central services (paid for out of general overheads) having no marginal cost.
10 comments
[ 3.3 ms ] story [ 30.6 ms ] threadhttps://doresearch.stanford.edu/research-scholarship/computi...
Answering that requires computing the 361st power of a very sparse 363 billion by 363 billion matrix, a big computation indeed (see http://www.cwi.nl/~tromp/go/legal.html for details).
Right now, I'm computing the number of 18x18 positions (standard size is 19x19) using a couple of servers with 512GB memory and a few TB of disk-space, and that's taking almost a year.
I'd love to have access to a super-computer, but they tend to frown upon such frivolous computations...
I know simple induction will not work for this problem, but to me, this feels like the kind of problem that might be simple enough that some mathematical breakthrough could make it a lot simpler.
I agree though, this does seem like the type of problem that should have a relatively simple solution, or at least a reasonable approximation.
The HPC does get used of course, but not for purely technical reasons. Genomics workloads should move to the cloud and be run at off-peak times to take advantage of cheaper pricing, but there are significant barriers to this as researchers can often get started on the HPC for free whereas cloud services require a purchasing card. Their cluster will also have some support and someone who can install software dependencies into the stack.
(As a result most genomics software is really, really hard to get running on anything other than the author's institutions HPC and its unique set of installed packages.)
Another barrier is just knowledge (it's hard to keep those people around in the current market) and university projects often have strange incentices due to central services (paid for out of general overheads) having no marginal cost.