Page 5 of 5 FirstFirst ... 345
Results 41 to 48 of 48

Thread: Chess Players! - A New Chess Engine!

  1. #41
    Join Date
    Jun 2008
    Location
    Podunk
    Posts
    27,088
    Blog Entries
    15

    Default Re: Chess Players! - A New Chess Engine!

    Hi
    TGIF

    So rebuilt the latest lc0 locally with cuda support... getting a better time...

    Code:
    lc0 benchmark --weights=/data/configuration/weights
           _
    |   _ | |
    |_ |_ |_| v0.23.0-dev+git.unknown built Oct  4 2019
    Loading weights file from: /data/configuration/weights
    Creating backend [cudnn]...
    GPU: GeForce GT 710
    GPU memory: 0.958313 Gb
    GPU clock frequency: 954 MHz
    GPU compute capability: 3.5
    CUDA Runtime version: 10.1.0
    Cudnn version: 7.6.3
    Latest version of CUDA supported by the driver: 10.1.0
    Benchmark time 607ms, 2 nodes, 3 nps, move e2e4
    .....
    Benchmark time 9106ms, 359 nodes, 39 nps, move e2e4
    bestmove e2e4
    Benchmark final time 9.79485s calculating 40.2252 nodes per second.
    Cheers Malcolm °¿° SUSE Knowledge Partner (Linux Counter #276890)
    SUSE SLE, openSUSE Leap/Tumbleweed (x86_64) | GNOME DE
    If you find this post helpful and are logged into the web interface,
    please show your appreciation and click on the star below... Thanks!

  2. #42
    Join Date
    Jun 2008
    Location
    San Diego, Ca, USA
    Posts
    11,371
    Blog Entries
    2

    Default Re: Chess Players! - A New Chess Engine!

    Thx Malcolm!

    And, just a general FYI to everyone, the LcO engine has proven not to be a "one shot wonder," it's pumping out consistently high quality play in the various competitions on TCEC and CCC. The displaced champion Stockfish (latest available is always available in the openSUSE repos) has been ramping up its development and pretty much is neck and neck with Lc0 with most of the other competitors falling back. An interesting new upstart is Leelenstein, which is based on Lc0 but with its author's custom tweaks. Sometimes it's better than both Lc0 and Stockfish, sometimes it falls back a bit. There are many semi-clones of Lc0 now including Leelenstein because Lc0 takes a hard line on excluding any kind of human input into its training (only tweaks to how Lc0 learns, but never the learning directly), the semi-clones are willing to break that rule and tell their engines how they think play should be improved.

    Whether you're a fan of 100% machine self-learning (Lc0) or the idea that humans can still write better software than a machine (Stockfish), both these applications are great chess programs, world class. And, both are FOSS.

    TSU
    Beginner Wiki Quickstart - https://en.opensuse.org/User:Tsu2/Quickstart_Wiki
    Solved a problem recently? Create a wiki page for future personal reference!
    Learn something new?
    Attended a computing event?
    Post and Share!

  3. #43
    Join Date
    Mar 2008
    Location
    Phuket, Thailand
    Posts
    26,569
    Blog Entries
    38

    Default Re: Chess Players! - A New Chess Engine!

    Quote Originally Posted by tsu2 View Post
    Whether you're a fan of 100% machine self-learning (Lc0) or the idea that humans can still write better software than a machine (Stockfish), both these applications are great chess programs, world class. And, both are FOSS.
    I agree fully its great to have both Lc0 and Stockfish available for GNU/Linux as FOSS, and indeed for openSUSE.

    I've played around with both, and while I am very happy to see Lc0, my experience with Lc0 on my desktop is that the Packman packaged stockfish20 is much stronger than the latest packaged Lc0 ... and I don't know why that is.

    For tuning Lc0 I took what looked like the strongest "weights" file, and also an opening database, and then had it play stockfish20, giving Lc0 15-minutes for all of its moves, and giving stockfish20 only 5-minutes for all of its moves (ie Lc0 had 3x as much time). After 6 games, I judged 3 games as ties (where stockfish's slight material edge was not enough to win after 80 some moves), and stockfish won the other 3 games all in less than 45 moves (delivering checkmate). At no time did Lc0 have a winning advantage in any of the 6-games.

    This is inconsistent with the results we see in the internet setup tournaments.

    And, ok , 6 is a small number of games. Likely there is some 'custom tuning' needed for my hardware.

    And further, the packaged (for openSUSE) Lc0 (even if weaker than the packaged stockfish20) still plays a very strong game of chess - far better than what I can play.

    .
    Its fun playing against both.
    .

  4. #44
    Join Date
    Jun 2008
    Location
    Podunk
    Posts
    27,088
    Blog Entries
    15

    Default Re: Chess Players! - A New Chess Engine!

    Quote Originally Posted by oldcpu View Post
    I agree fully its great to have both Lc0 and Stockfish available for GNU/Linux as FOSS, and indeed for openSUSE.

    I've played around with both, and while I am very happy to see Lc0, my experience with Lc0 on my desktop is that the Packman packaged stockfish20 is much stronger than the latest packaged Lc0 ... and I don't know why that is.

    For tuning Lc0 I took what looked like the strongest "weights" file, and also an opening database, and then had it play stockfish20, giving Lc0 15-minutes for all of its moves, and giving stockfish20 only 5-minutes for all of its moves (ie Lc0 had 3x as much time). After 6 games, I judged 3 games as ties (where stockfish's slight material edge was not enough to win after 80 some moves), and stockfish won the other 3 games all in less than 45 moves (delivering checkmate). At no time did Lc0 have a winning advantage in any of the 6-games.

    This is inconsistent with the results we see in the internet setup tournaments.

    And, ok , 6 is a small number of games. Likely there is some 'custom tuning' needed for my hardware.

    And further, the packaged (for openSUSE) Lc0 (even if weaker than the packaged stockfish20) still plays a very strong game of chess - far better than what I can play.

    .
    Its fun playing against both.
    .
    Hi
    Can you provide a link to the weights file you used? It might be just picking the right backend combination. Do you nvidia hardware?
    Cheers Malcolm °¿° SUSE Knowledge Partner (Linux Counter #276890)
    SUSE SLE, openSUSE Leap/Tumbleweed (x86_64) | GNOME DE
    If you find this post helpful and are logged into the web interface,
    please show your appreciation and click on the star below... Thanks!

  5. #45
    Join Date
    Mar 2008
    Location
    Phuket, Thailand
    Posts
    26,569
    Blog Entries
    38

    Default Re: Chess Players! - A New Chess Engine!

    Quote Originally Posted by malcolmlewis View Post
    Can you provide a link to the weights file you used? It might be just picking the right backend combination. Do you nvidia hardware?
    I have only intel hardware on an old core-i7. An extract from "inxi -F"
    Code:
    System:    Host: linux-p15v Kernel: 4.12.14-lp150.12.73-default x86_64 bits: 64 Desktop: KDE Plasma 5.12.8
               Distro: openSUSE Leap 15.0
    Machine:   Device: desktop System: Gigabyte product: Z87X-D3H serial: N/A
               Mobo: Gigabyte model: Z87X-D3H-CF v: x.x serial: N/A UEFI: American Megatrends v: F7 date: 08/02/2013
    CPU:       Quad core Intel Core i7-4770 (-HT-MCP-) cache: 8192 KB
               clock speeds: max: 3900 MHz 1: 3392 MHz 2: 3392 MHz 3: 3392 MHz 4: 3392 MHz 5: 3392 MHz 6: 3392 MHz
               7: 3392 MHz 8: 3392 MHz
    Graphics:  Card: Intel Xeon E3-1200 v3/4th Gen Core Processor Integrated Graphics Controller
               Display Server: x11 (X.Org 1.19.6 ) drivers: modesetting (unloaded: fbdev,vesa)
               Resolution: 1920x1200@59.95hz
               OpenGL: renderer: Mesa DRI Intel Haswell Desktop version: 4.5 Mesa 18.0.2
    After reading your post, I went to LCZ rating list (CCRL estimate), and selected "Lc0 ID 42700" from that list, and then went to Lc0 networks and looked for 42700 which gave "03c8b3db" for download. I downloaded that weighting file, and tried it. I noted it is much larger in file size (52.9 MB) than the previous weighting files (6.2 MB) I had used.

    Lc0 played a stronger game with that "Lc0 ID 42700" and won a couple of games vs stockfish20, and also lost a couple of games vs stockfish20. So it was stronger on my hardware and maybe equal to stockfish20 .... almost .... as I was giving stockfish only 1/3 the time I was giving Lc0.

  6. #46
    Join Date
    Jun 2008
    Location
    Podunk
    Posts
    27,088
    Blog Entries
    15

    Default Re: Chess Players! - A New Chess Engine!

    Quote Originally Posted by oldcpu View Post
    I have only intel hardware on an old core-i7. An extract from "inxi -F"
    Code:
    System:    Host: linux-p15v Kernel: 4.12.14-lp150.12.73-default x86_64 bits: 64 Desktop: KDE Plasma 5.12.8
               Distro: openSUSE Leap 15.0
    Machine:   Device: desktop System: Gigabyte product: Z87X-D3H serial: N/A
               Mobo: Gigabyte model: Z87X-D3H-CF v: x.x serial: N/A UEFI: American Megatrends v: F7 date: 08/02/2013
    CPU:       Quad core Intel Core i7-4770 (-HT-MCP-) cache: 8192 KB
               clock speeds: max: 3900 MHz 1: 3392 MHz 2: 3392 MHz 3: 3392 MHz 4: 3392 MHz 5: 3392 MHz 6: 3392 MHz
               7: 3392 MHz 8: 3392 MHz
    Graphics:  Card: Intel Xeon E3-1200 v3/4th Gen Core Processor Integrated Graphics Controller
               Display Server: x11 (X.Org 1.19.6 ) drivers: modesetting (unloaded: fbdev,vesa)
               Resolution: 1920x1200@59.95hz
               OpenGL: renderer: Mesa DRI Intel Haswell Desktop version: 4.5 Mesa 18.0.2
    After reading your post, I went to LCZ rating list (CCRL estimate), and selected "Lc0 ID 42700" from that list, and then went to Lc0 networks and looked for 42700 which gave "03c8b3db" for download. I downloaded that weighting file, and tried it. I noted it is much larger in file size (52.9 MB) than the previous weighting files (6.2 MB) I had used.

    Lc0 played a stronger game with that "Lc0 ID 42700" and won a couple of games vs stockfish20, and also lost a couple of games vs stockfish20. So it was stronger on my hardware and maybe equal to stockfish20 .... almost .... as I was giving stockfish only 1/3 the time I was giving Lc0.
    Hi
    Look at configuring it to use demux and 4 cores? Try the following;

    Code:
    lc0 benchmark --weights=/data/configuration/oldcpu -b demux --backend-opts=backend=blas,a,b,c,d
           _
    |   _ | |
    |_ |_ |_| v0.23.0-dev+git.unknown built Oct  4 2019
    Loading weights file from: /data/configuration/oldcpu
    Creating backend [demux]...
    Creating backend [blas]...
    BLAS vendor: OpenBLAS.
    OpenBLAS [OpenBLAS 0.3.7 DYNAMIC_ARCH NO_AFFINITY Sandybridge MAX_THREADS=64].
    OpenBLAS found 8 Sandybridge core(s).
    OpenBLAS using 1 core(s) for this backend.
    BLAS max batch size is 256.
    Creating backend [blas]...
    ....
    Benchmark time 385ms, 4 nodes, 10 nps, move e2e4
    Benchmark time 773ms, 7 nodes, 9 nps, move e2e4
    ....
    Benchmark time 9507ms, 479 nodes, 50 nps, move e2e4
    bestmove e2e4
    Benchmark final time 10.3317s calculating 50.6207 nodes per second.
    Cheers Malcolm °¿° SUSE Knowledge Partner (Linux Counter #276890)
    SUSE SLE, openSUSE Leap/Tumbleweed (x86_64) | GNOME DE
    If you find this post helpful and are logged into the web interface,
    please show your appreciation and click on the star below... Thanks!

  7. #47
    Join Date
    Mar 2008
    Location
    Phuket, Thailand
    Posts
    26,569
    Blog Entries
    38

    Default Re: Chess Players! - A New Chess Engine!

    Quote Originally Posted by malcolmlewis View Post
    Hi
    Look at configuring it to use demux and 4 cores? Try the following;

    Code:
    lc0 benchmark --weights=/data/configuration/oldcpu -b demux --backend-opts=backend=blas,a,b,c,d
    When I run 'top' during its execution, it is using only 1-core. Stockfish20, on the other hand, is using 2-cores.
    Code:
    oldcpu@linux-p15v:~/rpms-programs/arenalinux_64bit_1.1/Engines/lc0> lc0 benchmark --weights=/home/oldcpu/rpms-programs/arenalinux_64bit_1.1/Engines/lc0/weights.bin -b demux --backend-opts=backend=blas,a,b,c,d
           _
    |   _ | |
    |_ |_ |_| v0.23.0-dev+git.unknown built Sep 18 2019
    Loading weights file from: /home/oldcpu/rpms-programs/arenalinux_64bit_1.1/Engines/lc0/weights.bin
    Creating backend [demux]...
    Creating backend [blas]...
    BLAS vendor: OpenBLAS.
    OpenBLAS [DYNAMIC_ARCH Haswell].
    OpenBLAS found 4 Haswell core(s).
    OpenBLAS using 1 core(s) for this backend.
    BLAS max batch size is 256.
    Creating backend [blas]...
    BLAS vendor: OpenBLAS.
    OpenBLAS [DYNAMIC_ARCH Haswell].
    OpenBLAS found 4 Haswell core(s).
    OpenBLAS using 1 core(s) for this backend.
    BLAS max batch size is 256.
    Creating backend [blas]...
    BLAS vendor: OpenBLAS.
    OpenBLAS [DYNAMIC_ARCH Haswell].
    OpenBLAS found 4 Haswell core(s).
    OpenBLAS using 1 core(s) for this backend.
    BLAS max batch size is 256.
    Creating backend [blas]...
    BLAS vendor: OpenBLAS.
    OpenBLAS [DYNAMIC_ARCH Haswell].
    OpenBLAS found 4 Haswell core(s).
    OpenBLAS using 1 core(s) for this backend.
    BLAS max batch size is 256.
    Benchmark time 288ms, 4 nodes, 13 nps, move e2e4
    Benchmark time 568ms, 7 nodes, 12 nps, move e2e4
    ...
    Benchmark time 9507ms, 852 nodes, 89 nps, move e2e4
    bestmove e2e4
    Benchmark final time 10.0368s calculating 89.2719 nodes per second.

  8. #48
    Join Date
    Jun 2008
    Location
    San Diego, Ca, USA
    Posts
    11,371
    Blog Entries
    2

    Default Re: Chess Players! - A New Chess Engine!

    There is an online Lc0 competitor using only the CPU and no GPU I've seen competing recently on CCC called Lc0-CPU.
    It does perform at a lesser level, below Stockfish and the Lc0 using GPUs but is generally near the top of the pack of other CPU only and newer GPU engines.

    Although there is no way to to draw direct comparisons between these online competitors which are generally running on at least moderately high end hardware (current hardware used for CCC tournaments generally run in the $3000-5000 range which is nothing compared to the recent Finals which ran on $25,000 hardware) compared to a home system which may not exceed $1500-3000,

    Most modern systems today are at least quad core (although if your system is not brand new may be dual), so the number of available "cpus" which also mean cores should be satisfactory but you may want to try forcing a different number of CPUs than default. I was informed awhile back though that running only on the CPU does not benefit greatly by increased CPUs and might actually deteriorate (of course has to do with task allocation efficiency)
    GPU engines tend to look at only about 1/10 as many positions as CPU engines, but I assume much more computation goes into evaluating/pruning the positions.
    CPU engines tend to benefit more by longer time controls when playing against GPU engines. The CCC online games are generally 10 minutes on a side plus 3 or 5 seconds per move, but the performance I'm describing is based on the hardware they're running.
    Perhaps most surprisingly, "stronger" nets don't always produce better results, for whatever reason sometimes they frequently under-perform using different time controls(sometimes longer, sometimes shorter, there isn't usually a reliable pattern).

    The problem with NN is that it's nearly impossible for humans to dig into a net and make sense of it, so one of the lessons for AI based on NN in general but especially when the engine is 100% self-learning is that we can never know how the engine taught itself and never know by what logic it "thinks" so when we believe that a machine's decision is "superior" or "high quality" we say that in good faith and not because we have any inkling whether it's truly good or bad. All we have to go on is that with more time given the machine to self-learn that it should be getting better, and how it might fare in test battles with known benchmarks like a certain version of Stockfish or a previous version of Lc0. But otherwise, it's not like we can say "This net is better because it now understands this concept better than before."

    HTH,
    TSU
    Beginner Wiki Quickstart - https://en.opensuse.org/User:Tsu2/Quickstart_Wiki
    Solved a problem recently? Create a wiki page for future personal reference!
    Learn something new?
    Attended a computing event?
    Post and Share!

Page 5 of 5 FirstFirst ... 345

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •