CMCanavessi wrote:Well, finally the 128 network has been made public. I'm broadcasting right now a 10-game match between LCZero ID 123 Vs. Fruit 1.0
Time control is 40/40, 5-move book with reverse games and Fruit has ponder on (which LCZero doesn't support). Even with that advantage in favor of Fruit, Leela is winning 1-0 and about to win the reverse game also, which would be 2-0.
You look like an insider there, so you might know about some upcoming changes. Do you know when the the v0.5 client will be updated? It seems buggy is some respects, for example "go infinite" doesn't work, which impedes me to test on test-suites in Polyglot.
Then, as my GT 730 GPU is an utter crap, with the new 128x10 networks CPU speed gets hit hard. A factor of 2 or so at its peak on initial positions (from 2000 NPS to 1000 or so). But even worse at short time controls I was testing before, maybe a factor of 4-5. Look at this progression in NPS with the new net:
NPS maxes at depth 27 after a whopping 4 minutes on 4 threads (i7 Haswell 3.8 GHz). And the increase in NPS is 3-fold from 1-3 seconds per move to 4 minutes per move. I will have some difficulty checking for "absolute" (say CCRL) strength without a very strong GPU, I hope the progress will be visible even in short tc games on CPU, although "absolute" ratings will be way off (by hundreds of Elo points).
Laskos wrote:
NPS maxes at depth 27 after a whopping 4 minutes on 4 threads (i7 Haswell 3.8 GHz). And the increase in NPS is 3-fold from 1-3 seconds per move to 4 minutes per move. I will have some difficulty checking for "absolute" (say CCRL) strength without a very strong GPU, I hope the progress will be visible even in short tc games on CPU, although "absolute" ratings will be way off (by hundreds of Elo points).
I am still tracking the progress with my meager CPU means. At 1s/move, I took Jabba 1.0 (about 2050 Elo CCRL) reference in matches of 200 games. First and second "bignets" performed as:
ID123: 57.5/200
ID124: 65.0/200
Although still within error margins, it seems it progresses fast, about +30 Elo points from one net to the successive one. I will maybe do a larger test when more progress accumulates (after several nets).
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
Laskos wrote:
NPS maxes at depth 27 after a whopping 4 minutes on 4 threads (i7 Haswell 3.8 GHz). And the increase in NPS is 3-fold from 1-3 seconds per move to 4 minutes per move. I will have some difficulty checking for "absolute" (say CCRL) strength without a very strong GPU, I hope the progress will be visible even in short tc games on CPU, although "absolute" ratings will be way off (by hundreds of Elo points).
I am still tracking the progress with my meager CPU means. At 1s/move, I took Jabba 1.0 (about 2050 Elo CCRL) reference in matches of 200 games. First and second "bignets" performed as:
ID123: 57.5/200
ID124: 65.0/200
Although still within error margins, it seems it progresses fast, about +30 Elo points from one net to the successive one. I will maybe do a larger test when more progress accumulates (after several nets).
I'm currently testing against Colossus 2008b - CCRL 2642
I'm using Noomen short and 15 seconds per move each.
Very early days, but current result for LCZero 127 on NVidia 1060 vs Colossus 2008b @ 15 sec / move below. Colossus is single Intel Broadwell core @ 4.2 GHz.
CMCanavessi wrote:No progress? We have just tested a new bigger net (128x10) and it's 200+ elo stronger.
What kind of filters are these networks using? Is that 3x3 convolutions like AlphaZero? These might be very good for Go, but for Chess, where there are sliders, they are blind to a lot of elementary patterns, such as pins and discovered threats. They would have to make up for that by cascading many blocks to even get the whole significant part 'in view' (e.g. a Rook on e1 pinning a Queen on e4 against a King on e8).
I wonder what would happen if you used filters that each examined one file, one rank and one diagonal as well. These would only look at 8 squares, instead of 9, so it is not more complex. E.g. of every 8 filters you could put 1 on each ray, and leave 4 to examine the local 3x3 neighborhood. That might give you a much more powerful network for the same number of weights.
Werewolf wrote:Very early days, but current result for LCZero 127 on NVidia 1060 vs Colossus 2008b @ 15 sec / move below. Colossus is single Intel Broadwell core @ 4.2 GHz.
4 wins
2 losses
1 draw
for LCZero
Wow, on a good GPU and longer time control, LC0 rocks. It scales completely differently from standard engines, give it strong hardware and LTC, and it soares.