Merge pull request #301 from ujh/release-0.3.2

Release 0.3.2

CHANGELOG.md

@@ -1,3 +1,36 @@
+## 0.3.2 [☰](https://github.com/ujh/iomrascalai/compare/0.3.1...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 [☰](https://github.com/ujh/iomrascalai/compare/0.3.0...0.3.1)
 
 * Implemented `gogui-analyze_commands` and `imrscl-safe_territory`

Cargo.toml

@@ -2,7 +2,7 @@
 
 description = "An AI for the game of Go"
 name = "iomrascalai"
-version = "0.3.1"
+version = "0.3.2"
 authors = ["Urban Hafner <[email protected]>",
         "Thomas Poinsot <[email protected]>",
         "Igor Polyakov <[email protected]>"]

bin/wins-from-benchmark-results.rb

@@ -40,7 +40,7 @@ def wins(fn)
   black = data.find_all {|row| row[RES_R] =~ /B\+/ }.count
   n = white + black
   p = white.to_f/n
-  "#{(p*100).round(2)}% wins (#{white} games of #{n}, ± #{error(p: p, n: n, confidence: 0.95).round(2)} at 95%, ± #{error(p: p, n: n, confidence: 0.99).round(2)} at 99%)"
+  "#{(p*100).round(2)}% wins (#{white} games of #{n}, ± #{error(p, n, 0.95).round(2)} at 95%, ± #{error(p, n, 0.99).round(2)} at 99%)"
 end
 
 def scoring(fn)
@@ -49,16 +49,16 @@ def scoring(fn)
   agreeing = relevant.find_all {|row| row[RES_W] == row[RES_B] }.count
   n = relevant.length
   p = agreeing.to_f/n
-  "#{(p*100).round(2)}% same score as GnuGo (#{agreeing} of #{n}, ± #{error(p: p, n: n, confidence: 0.95).round(2)} at 95%, ± #{error(p: p, n: n, confidence: 0.99).round(2)} at 99%)"
+  "#{(p*100).round(2)}% same score as GnuGo (#{agreeing} of #{n}, ± #{error(p, n, 0.95).round(2)} at 95%, ± #{error(p, n, 0.99).round(2)} at 99%)"
 end
 
-def z(confidence:)
+def z(confidence)
   alpha = 1 - confidence
   (1 - 0.5*alpha)*2
 end
 
-def error(p:, n:, confidence:)
-  (z(confidence: confidence) * Math.sqrt((1.0/n)*p*(1-p)))*100
+def error(p, n, confidence)
+  (z(confidence) * Math.sqrt((1.0/n)*p*(1-p)))*100
 end
 
 Dir["*.dat"].each do |fn|

src/version.rs

@@ -1,6 +1,7 @@
 /************************************************************************
  *                                                                      *
  * Copyright 2015 Urban Hafner                                          *
+ * Copyright 2016 Urban Hafner                                          *
  *                                                                      *
  * This file is part of Iomrascálaí.                                    *
  *                                                                      *
@@ -20,5 +21,5 @@
  ************************************************************************/
 
 pub fn version() -> &'static str {
-    "0.3.1"
+    "0.3.2"
 }