[Pachi] Question about UCT tree structure

Stanislaw Halik sthalik at misaki.pl
Fri May 13 05:09:58 CEST 2016


On 2016-04-22 09:43, Jonathan Ho wrote:
> It seems that all of the UCT code assumes that this board stays in
> memory during the entire duration of existence of the uct engine
> (from engine_uct_init until uct_done), because the tree is stored
> inside the uct data structure.

At the very least, "struct tree_node" is very sensitive to cache misses.
Removing 10% worth of only padding yielded about 10% improvement in
evaluations/s.

There are probably logic errors if "struct board" isn't kept consistent,
pasky or someone else can elaborate on the logic. I can't.

 > I'm interested in writing Python wrappers for the various primitives
 > in Pachi, such as boards and engines. In particular, I'm trying to
 > make the UCT code accessible as a library, so that a user can give
 > the UCT enginea board and ask it for a move.

> What is the meaning of this particular board? Is it safe to modify
> the tree structure so that it instead keeps a copy of a board instead
> of a pointer to it (i.e. modify tree_init to copy the board
> argument)?

I don't see a reason why to do that in the first place, taking account
Python's FFI approaches. At least please use cffi or other modern 
library, not direct cpython integration.

> If not, does this mean that with a uct engine, I must keep track of
> an associated board that it operates on, and ensure that it is never
> deallocated over the course of the engine's lifetime?

Pachi takes care of the lifetime. See the call graph, read the functions 
and you'll note it's not harder than for other C libraries.

If you need to marshal C <-> Python at all in the evaluation function, 
it's a big performance hit. Try to avoid that.

sh


More information about the Pachi mailing list