Permalink
Fetching contributors…
Cannot retrieve contributors at this time
263 lines (192 sloc) 9.93 KB

0.3.2

  • Fixes a bug where the bot passed too early and lost won games because of it. #261
  • Resign lost games earlier. This speeds up the benchmark and is also nicer for humans. #264
  • Speed up the calculations of the priors by moving them into the worker threads. #275
  • Implement the reg_genmove GTP command. #289
  • Implement the kgs-genmove_cleanup GTP command. #290
  • Improve the scoring accuracy when the game is over. #292
  • Fix bug where some rotations of the small patterns were missing. #299
  • Implemented the imrscl-uct_gfx custom GTP command. A GoGui analyze compatible command to show the number of playouts that went through each child node of the root (i.e. each point on the board) #293

Performance

The benchmarks were run on a c4.8xlarge EC2 instance with 8 worker threads.

After running 500 games on 9x9 with komi 6.5 and a time limit of 2 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 76.2% (± 3.71 at 95%, ± 3.79 at 99%) with the default configuration. It agreed on the final score with GnuGo in 74.09% (± 4.35 at 95%, ± 4.44 at 99%) of games.

After running 1000 games on 13x13 with komi 6.5 and a time limit of 10 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 42.8% (± 3.05 at 95%, ± 3.11 at 99%) with the default configuration. It agreed on the final score with GnuGo in 59.78% (± 4.53 at 95%, ± 4.63 at 99%) of games.

0.3.1

  • Implemented gogui-analyze_commands and imrscl-safe_territory which shows the intersections on the board that the bot thinks are safe (i.e. definitely belong to one color).
  • Stop playing when all intersections are decided (i.e. the bot thinks they belong to a certain color).
  • Implement final_status_list
  • Implement Chinese scoring
  • Return to command line arguments for turning on/off logging, GoGui effects, and ruleset configuration.

Performance

Note that we now use Chinese rules instead of CGOS/Tromp Taylor rules which may change the results.

After running 500 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 77.4% (± 3.65 at 95%, ± 3.72 at 99%) with the default configuration.

After running 500 games on 13x13 with komi 6.5 and a time limit of 10 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 35.2% (± 4.16 at 95%, ± 4.25 at 99%) with the default configuration.

0.3.0

  • Added 3x3 pattern support. These are used as priors in the tree and during the playouts.
  • Added the RAVE heuristic to the search.
  • Terminate search early if the best move is ahead by a lot.
  • Changed the way to change the configuration of the engine. Everything is now stored inside a TOML file instead of having command line switches for everything.
  • Measure the playouts per second per thread and display it when logging is turned on.
  • Scripts to run CLOP parameter optimization.
  • Removed all weak engines (random, AMAF, MC).
  • Clarified the licensing (scripts are MIT licensed).
  • Added a code of conduct.

Performance

Note that the measurement of the error margins have changed. We now calculate proper 95% and 99% confidence intervals whereas previously we only used the ~68% standard deviation.

After running 500 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 85.40% (± 3.08 at 95%, ± 3.14 at 99%) with the default configuration.

After running 500 games on 13x13 with komi 6.5 and a time limit of 10 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 36.6% (± 4.2 at 95%, ± 4.29 at 99%) with the default configuration.

0.2.4

This release doesn't contain any new features. But as the last release happened quite a long time ago and there were some (small) internal changes a new release seems to be in order. Also the Rust compiler changed a lot in between and changed the performance.

Performance

After running 200 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 65.5% ± 3.4 for the default engine with 8 threads.

After running 200 games on 13x13 with komi 6.5 and a time limit of 10 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 10% ± 2.1 for the default engine with 8 threads.

0.2.3

  • Playout the aftermath by default on CGOS
  • Play in the middle of big eyes during playouts
  • Resign lost games faster (to save time)

Performance

After running 200 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 75.5% ± 3% for the default engine with 8 threads.

After running 200 games on 13x13 with komi 6.5 and a time limit of 10 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 6% ± 1.7% for the default engine with 8 threads.

0.2.2

  • A two liberty solver (can prove if a group with up to two stones is dead)
  • A set of priors for the UCT tree
    • One that prioritizes killing opponent groups
    • One that discourages putting one's own groups into danger
    • One to disourage plays on the first two lines if no other stones are close by

Performance

After running 200 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 70% ± 3.2% for the default engine with 8 threads.

After running 200 games on 13x13 with komi 6.5 and a time limit of 10 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 11% ± 2.2% for the default engine with 8 threads.

0.2.1

  • New UCT policy UCB1tuned
  • Reuse the subtree from the previous genmove
  • Seki detection in the self atari code in the playouts

Performance

After running 100 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rate against GnuGo 3.8 level 0 was 35% ± 4.8% for the default engine with 8 threads.

0.2.0

  • UCT engine
  • A new playout policy that doesn't put large strings into self atari

0.1.7

  • Improvements of the playout speeds
  • The playout type can now be selected from the command line (use the -p command line option)

0.1.6

  • The engines can now utilize multiple threads (use the -t command line switch)
  • Debug output to stderr is now only generated when running with -l
  • The code coverage is now measured and reported through a badge in the README

0.1.5

Changes

  • Fix bug in AMAF code that made it much weaker than it should be. The first move of each playout wasn't taken into account during the move selection (#102)
  • Implemented the loadsgf GTP command (#104)

Performance

After running 100 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rates are as follows:

  • 3% ± 1.7% of the MC engine against GnuGo 3.8 (level 0)
  • 3% ± 1.7% of the AMAF engine against GnuGo 3.8 (level 0)
  • 92% ± 2.7% of the MC engine against AmiGoGtp 1.8
  • 92% ± 2.7% of the AMAF engine against AmiGoGtp 1.8
  • 41% ± 4.9% of the MC engine against the MC engine of 0.1.4
  • 52% ± 5% of the AMAF engine against the AMAF engine of 0.1.4
  • 70% ± 4.6% of the current AMAF engine against the current MC engine

0.1.4

Changes

  • Make AMAF the default engine. This was supposed to be the default for 0.1.3, but didn't happen so the win rates recorded for 0.1.3 are useless!
  • Refactored the engine code, to but the shared code of the MC and the AMAF into a trait.

Performance

After running 100 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rates are as follows:

  • 10% ± 3% against Iomrascálaí 0.1.3
  • 2% ± 1.4% against GnuGo 3.8 (Level 0)
  • 70% ± 4.6% win rate of the AMAF engine against the MC engine

0.1.3

Changes

  • A new engine implementing the AMAF (all moves as first) pattern. Almost the same as standard Monte-Carlo, but recording wins and losses for all moves of the simulated color.
  • Don't pass even when the win rate calculated through the simulations is 100%. Doing this resulted in losses against a random player.
  • Scripts to play a game in GoGui against GnuGo and Brown.
  • Refactored the time keeping code. It has now moved out of the engines and lives in a separate struct.

Performance

After running 100 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rates are as follows:

  • 49% ± 5% against Iomrascálaí 0.1.2
  • 5% ± 2.2% against GnuGo 3.8 (Level 0)

0.1.2

Changes

  • A small change to the playout policy: Only play the pass move if no other moves (that are not plays into single point eyes) are possible.
  • Add GnuGo as the referee in the benchmark, so that we always get a score, even if the bots disagree.
  • Enhance the benchmark script to also play against the previous version of the bot.

Performance

After running 100 games on 9x9 with komi 6.5 and a time limit of 5 minutes (sudden death) the win rates are as follows:

  • 1% ± 1% against Iomrascálaí 0.1.1
  • 7% ± 2.6% against GnuGo 3.8 (Level 0)