4 # Author:: Daigo Moriwaki
5 # Homepage:: http://sourceforge.jp/projects/shogi-server/
8 # Copyright (C) 2006-2012 Daigo Moriwaki <daigo at debian dot org>
10 # This program is free software; you can redistribute it and/or modify
11 # it under the terms of the GNU General Public License as published by
12 # the Free Software Foundation; either version 2 of the License, or
13 # (at your option) any later version.
15 # This program is distributed in the hope that it will be useful,
16 # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 # GNU General Public License for more details.
20 # You should have received a copy of the GNU General Public License
21 # along with this program; if not, write to the Free Software
22 # Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
27 # mk_rate reads game results files generated by the mk_game_results command,
28 # calculates rating scores of each player, and then outputs a yaml file
29 # (players.yaml) that Shogi-server can recognize.
33 # ./mk_rate [options] GAME_RESULTS_FILE [...]
38 # a path to a file listing results of games, which is generated by the
39 # mk_game_results command.
40 # In the second style above, the file content can be read from the stdin.
42 # --abnormal-threshold::
43 # n [plies] (default 30)
44 # Games that end with the 'abnormal' status are counted in win/lost games
45 # for the rating calculation if a game plays more than n plies. Otherwise
46 # (or if n is zero), abnormal games are counted out of rating games.
49 # a base time point for this calculation (default now). Ex. '2009-10-31'
52 # n [days] (default 60)
54 # --half-life-ignore::
55 # m [days] (default 7)
56 # after m days, the half-life effect works
59 # m [days] (default 365*2)
60 # old results will be ignored
62 # --fixed-rate-player::
63 # player whose rate is fixed at the rate
69 # skip draw games. [default: draw games are counted in as 0.5 win and 0.5
77 # Sample Command lines that install prerequires will work on Debian.
79 # * Ruby 2.0.0 or later (including Rubygems)
81 # $ sudo aptitude install ruby
83 # * Ruby bindings for the GNU Scientific Library (GSL[http://rb-gsl.rubyforge.org/])
85 # $ sudo aptitude install ruby-gsl
87 # * RGL: {Ruby Graph Library}[http://rubyforge.org/projects/rgl/]
89 # $ sudo gem install rgl
93 # $ ./mk_rate game_results.txt > players.yaml
95 # $ ./mk_game_results . | ./mk_rate > players.yaml
97 # If you do not want the file to be update in case of errors,
99 # $ ./mk_rate game_results.txt && ./mk_rate game_results.txt > players.yaml
101 # == How players are rated
103 # The conditions that games and players are rated as following:
105 # * Rated games, which were played by both rated players.
106 # * Rated players, who logged in the server with a name followed by a trip: "name,trip".
107 # * (Rated) players, who played more than $GAMES_LIMIT [15] (rated) games.
111 $:.unshift(File.dirname(Pathname.new(__FILE__).realpath))
112 require 'utils/csa-filter'
119 require 'rgl/adjacency'
120 require 'rgl/connected_components'
122 #################################################
126 # Count out players who play less games than $GAMES_LIMIT
127 $GAMES_LIMIT = $DEBUG ? 0 : 15
134 # Holds the last time when a player gamed
135 $players_time = Hash.new { Time.at(0) }
136 # Holds history of input lines to check duplicated inputs
140 #################################################
141 # Keeps the value of the lowest key
149 if @lowest.empty? || key < @lowest[0]
150 @lowest = [key, value]
163 #################################################
164 # Calculates rates of every player from a Win Loss GSL::Matrix
169 # The model of the win possibility is 1/(1 + 10^(-d/400)).
170 # The equation in this class is 1/(1 + e^(-Kd)).
171 # So, K should be calculated like this.
172 K = Math.log(10.0) / 400.0
174 # Convergence limit to stop Newton method.
176 # Stop Newton method after this iterations.
179 # Average rate among the players
188 # Calcurates the average of the vector.
190 def Rating.average(vector, mean=0.0)
191 sum = Array(vector).inject(0.0) {|sum, n| sum + n}
192 vector -= GSL::Vector[*Array.new(vector.size, sum/vector.size - mean)]
199 def initialize(win_loss_matrix)
203 when GSL::Matrix, GSL::Matrix::Int
212 attr_reader :rate, :n
216 (0...@size).collect {|k| yield k}
221 (0...@size).each {|k| yield k}
225 # The possibility that the player k will beet the player i.
228 1.0/(1.0 + exp(@rate[i]-@rate[k]))
232 # Most possible equation
239 sum += @n[k,i] * win_rate(i,k) - @n[i,k] * win_rate(k,i)
246 # / f0/R0 f0/R1 f0/R2 ... \
247 # dfk/dRj = | f1/R0 f1/R1 f1/R2 ... |
248 # \ f2/R0 f2/R1 f2/R2 ... /
254 sum += win_rate(i,k) * win_rate(k,i) * (@n[k,i] + @n[i,k])
258 sum = 2.0 * win_rate(j,k) * win_rate(k,j) * (@n[k,j] + @n[j,k])
264 # Jacobi matrix of the func().
270 (0...@size).collect do |k|
271 (0...@size).collect do |j|
279 # The initial value of the rate, which is of very importance for Newton
280 # method. This is based on my huristics; the higher the win probablity of
281 # a player is, the greater points he takes.
286 v = GSL::Vector[0, 0]
289 v += GSL::Vector[@n[k,i], @n[i,k]]
291 v.nrm2 < 1 ? 0 : v[0] / (v[0] + v[1])
293 rank = possibility.sort_index
294 @rate = player_vector do |k|
295 K*500 * (rank[k]+1) / @size
301 # Resets @rate as the higher the current win probablity of a player is,
302 # the greater points he takes.
305 @rate = @record.get || @rate
306 rank = @rate.sort_index
307 @rate = player_vector do |k|
308 K*@count*1.5 * (rank[k]+1) / @size
313 # mu is the deaccelrating parameter in Deaccelerated Newton method
314 def deaccelrate(mu, old_rate, a, old_f_nrm2)
315 @rate = old_rate - a * mu
316 if func_vector.nrm2 < (1 - mu / 4.0 ) * old_f_nrm2 then
320 @record.set(func_vector.nrm2, @rate)
324 $stderr.puts "mu: %f " % [mu] if $DEBUG
325 deaccelrate(mu*0.5, old_rate, a, old_f_nrm2)
329 # Main process to calculate ratings.
332 # Counter to stop the process.
333 # Calulation in Newton method may fall in an infinite loop
338 # Solve the equation:
340 # @rate_(n+1) = @rate_(n) - a
342 # f.nrm2 should approach to zero.
346 # $stderr.puts "j: %s" % [j.inspect] if $DEBUG
347 $stderr.puts "f: %s -> %f" % [f.to_a.inspect, f.nrm2] if $DEBUG
349 # GSL::Linalg::LU.solve or GSL::Linalg::HH.solve would be available instead.
350 #a = GSL::Linalg::HH.solve(j, f)
351 a, = GSL::MultiFit::linear(j, f)
352 a = self.class.average(a)
353 # $stderr.puts "a: %s -> %f" % [a.to_a.inspect, a.nrm2] if $DEBUG
355 # Deaccelerated Newton method
356 # GSL::Vector object should be immutable.
359 old_f_nrm2 = old_f.nrm2
360 deaccelrate(1.0, old_rate, a, old_f_nrm2)
361 #@rate -= a # Instead, do not deaccelerate
362 @record.set(func_vector.nrm2, @rate)
364 $stderr.printf "|error| : %5.2e\n", a.nrm2 if $DEBUG
367 if @count > COUNT_MAX
368 $stderr.puts "Values seem to oscillate. Stopped the process."
369 $stderr.puts "f: %s -> %f" % [func_vector.to_a.inspect, func_vector.nrm2]
373 end while (a.nrm2 > ERROR_LIMIT * @rate.nrm2)
376 $stderr.puts "resolved f: %s -> %f" %
377 [func_vector.to_a.inspect, func_vector.nrm2] if $DEBUG
378 $stderr.puts "Count: %d" % [@count] if $DEBUG
386 # Make the values of @rate finite.
389 @rate = @rate.collect do |a|
391 a.infinite? * AVERAGE_RATE * 100
399 # Flatten the values of @rate.
401 def average!(mean=0.0)
402 @rate = self.class.average(@rate, mean)
408 def translate!(value)
413 # Make the values of @rate integer.
416 @rate = @rate.collect do |a|
422 a.infinite? * AVERAGE_RATE * 100
428 #################################################
429 # Encapsulate a pair of keys and win loss matrix.
430 # - keys is an array of player IDs; [gps+123, foo+234, ...]
431 # - matrix holds games # where player i (row index) beats player j (column index).
432 # The row and column indexes match with the keys.
434 # This object should be immutable. If an internal state is being modified, a
435 # new object is always returned.
443 def self.mk_matrix(players)
444 keys = players.keys.sort
448 ((0...size).collect do |k|
450 p1_hash = players[p1]
451 ((0...size).collect do |j|
456 v = p1_hash[p2] || GSL::Vector[0,0]
461 return WinLossMatrix.new(keys, matrix)
464 def self.mk_win_loss_matrix(players)
465 obj = mk_matrix(players)
473 # an array of player IDs; [gps+123, foo+234, ...]
476 # matrix holds games # where player i (row index) beats player j (column index).
477 # The row and column indexes match with the keys.
480 def initialize(keys, matrix)
486 # Returns the size of the keys/matrix
497 # Removes players in a rows such as [1,3,5], and then returns a new
500 def delete_rows(rows)
501 rows = rows.sort.reverse
504 (0...size).each do |i|
505 next if rows.include?(i)
506 row = @matrix.row(i).clone
512 if copied_cols.size == 0
513 new_matrix = GSL::Matrix.new
515 new_matrix = GSL::Matrix[*copied_cols]
518 new_keys = @keys.clone
520 new_keys.delete_at(j)
523 return WinLossMatrix.new(new_keys, new_matrix)
527 # Removes players who do not pass a criteria to be rated, and returns a
531 $stderr.puts @keys.inspect if $DEBUG
532 $stderr.puts @matrix.inspect if $DEBUG
534 (0...size).each do |i|
539 if win < 1 || loss < 1 || win + loss < $GAMES_LIMIT
544 # The recursion ends if there is nothing to delete
545 return self if delete.empty?
547 new_obj = delete_rows(delete)
552 # Cuts self into connecting groups such as each player in a group has at least
553 # one game with other players in the group. Returns them as an array.
555 def connected_subsets
556 g = RGL::AdjacencyGraph.new
557 (0...size).each do |k|
558 (0...size).each do |i|
567 g.each_connected_component do |c|
570 new_keys << keys[v.to_s.to_i]
575 subsets = subsets.sort {|a,b| b.size <=> a.size}
577 result = subsets.collect do |keys|
580 ((0...keys.size).collect do |k|
581 p1 = @keys.index(keys[k])
582 ((0...keys.size).collect do |j|
586 p2 = @keys.index(keys[j])
591 WinLossMatrix.new(keys, matrix)
598 "size : #{@keys.size}" + "\n" +
599 @keys.inspect + "\n" +
606 #################################################
611 # After NHAFE_LIFE days value will get half.
612 # 0.693 is constant, where exp(0.693) ~ 0.5
614 if days < $options["half-life-ignore"]
617 Math::exp(-0.693/$options["half-life"]*(days-$options["half-life-ignore"]))
621 def _add_win_loss(winner, loser, time)
622 how_long_days = ($options["base-date"] - time)/(3600*24)
623 $players[winner] ||= Hash.new { GSL::Vector[0,0] }
624 $players[loser] ||= Hash.new { GSL::Vector[0,0] }
625 $players[winner][loser] += GSL::Vector[1.0*half_life(how_long_days),0]
626 $players[loser][winner] += GSL::Vector[0,1.0*half_life(how_long_days)]
629 def _add_draw(player1, player2, time)
630 how_long_days = ($options["base-date"] - time)/(3600*24)
631 $players[player1] ||= Hash.new { GSL::Vector[0,0] }
632 $players[player2] ||= Hash.new { GSL::Vector[0,0] }
633 $players[player1][player2] += GSL::Vector[0.5*half_life(how_long_days),0.5*half_life(how_long_days)]
634 $players[player2][player1] += GSL::Vector[0.5*half_life(how_long_days),0.5*half_life(how_long_days)]
637 def _add_time(player, time)
638 $players_time[player] = time if $players_time[player] < time
641 def add(black_mark, black_name, white_name, white_mark, time)
642 if black_mark == WIN_MARK && white_mark == LOSS_MARK
643 _add_win_loss(black_name, white_name, time)
644 elsif black_mark == LOSS_MARK && white_mark == WIN_MARK
645 _add_win_loss(white_name, black_name, time)
646 elsif black_mark == DRAW_MARK && white_mark == DRAW_MARK
647 if $options["skip-draw-games"]
650 _add_draw(black_name, white_name, time)
653 raise "Never reached!"
655 _add_time(black_name, time)
656 _add_time(white_name, time)
660 if /@NORATE\+/ =~ id # the player having @NORATE in the name should not be rated
663 id.gsub(/@.*?\+/,"+")
666 # Parse a game result line
669 if $history.include? line
670 $stderr.puts "[WARNING] Duplicated: #{line}"
675 time, state, black_mark, black_id, white_id, white_mark, file = line.split("\t")
676 unless time && state && black_mark && black_id &&
677 white_id && white_mark && file
678 $stderr.puts "Failed to parse the line : #{line}"
682 if state == "abnormal"
683 csa = CsaFileReader.new(file, "EUC-JP")
684 if $options["abnormal-threshold"] == 0 || csa.ply <= $options["abnormal-threshold"]
688 time = Time.parse(time)
689 return if $options["base-date"] < time
690 how_long_days = ($options["base-date"] - time)/(3600*24)
691 if (how_long_days > $options["ignore"])
695 black_id = identify_id(black_id)
696 white_id = identify_id(white_id)
698 if black_id && white_id && (black_id != white_id) &&
699 black_mark && white_mark
700 add(black_mark, black_id, white_id, white_mark, time)
705 yaml["players"].each do |group_key, group|
706 group.each do |player_key, player|
707 rate = player['rate']
709 if rate > 10000 || rate < -10000
719 USAGE: #{$0} [options] GAME_RESULTS_FILE [...]
723 a path to a file listing results of games, which is genrated by the
724 mk_game_results command.
725 In the second style above, the file content can be read from the stdin.
728 --base-date a base time point for this calicuration (default now). Ex. '2009-10-31'
729 --half-life n [days] (default 60)
730 --half-life-ignore m [days] (default 7)
731 after m days, half-life effect works
732 --ignore n [days] (default 730 [=365*2]).
733 Results older than n days from the 'base-date' are ignored.
734 --fixed-rate-player player whose rate is fixed at the rate
736 --skip-draw-games skip draw games. [default: draw games are counted in
737 as 0.5 win and 0.5 lost]
738 --help show this message
744 parser = GetoptLong.new(
745 ["--abnormal-threshold", GetoptLong::REQUIRED_ARGUMENT],
746 ["--base-date", GetoptLong::REQUIRED_ARGUMENT],
747 ["--half-life", GetoptLong::REQUIRED_ARGUMENT],
748 ["--half-life-ignore", GetoptLong::REQUIRED_ARGUMENT],
749 ["--help", "-h", GetoptLong::NO_ARGUMENT],
750 ["--ignore", GetoptLong::REQUIRED_ARGUMENT],
751 ["--fixed-rate-player", GetoptLong::REQUIRED_ARGUMENT],
752 ["--fixed-rate", GetoptLong::REQUIRED_ARGUMENT],
753 ["--skip-draw-games", GetoptLong::NO_ARGUMENT])
756 parser.each_option do |name, arg|
758 $options[name] = arg.dup
760 if ( $options["fixed-rate-player"] && !$options["fixed-rate"]) ||
761 (!$options["fixed-rate-player"] && $options["fixed-rate"]) ||
762 ( $options["fixed-rate-player"] && $options["fixed-rate"].to_i <= 0)
768 raise parser.error_message
774 if $options["base-date"]
775 $options["base-date"] = Time::parse $options["base-date"]
777 $options["base-date"] = Time.now
779 $options["abnormal-threshold"] ||= 30
780 $options["abnormal-threshold"] = $options["abnormal-threshold"].to_i
781 $options["half-life"] ||= 60
782 $options["half-life"] = $options["half-life"].to_i
783 $options["half-life-ignore"] ||= 7
784 $options["half-life-ignore"] = $options["half-life-ignore"].to_i
785 $options["ignore"] ||= 365*2
786 $options["ignore"] = $options["ignore"].to_i
787 $options["fixed-rate"] = $options["fixed-rate"].to_i if $options["fixed-rate"]
790 while line = $stdin.gets do
794 while file = ARGV.shift do
795 File.open(file) do |f|
796 f.each_line do |line|
807 obj = WinLossMatrix::mk_win_loss_matrix($players)
808 obj.connected_subsets.each do |win_loss_matrix|
809 yaml["players"][rating_group] = {}
811 rating = Rating.new(win_loss_matrix.matrix)
813 rating.average!(Rating::AVERAGE_RATE)
816 if $options["fixed-rate-player"]
817 # first, try exact match
818 index = win_loss_matrix.keys.index($options["fixed-rate-player"])
819 # second, try regular match
821 win_loss_matrix.keys.each_with_index do |p, i|
822 if %r!#{$options["fixed-rate-player"]}! =~ p
828 the_rate = rating.rate[index]
829 rating.translate!($options["fixed-rate"] - the_rate)
833 win_loss_matrix.keys.each_with_index do |p, i| # player_id, index#
834 win = win_loss_matrix.matrix.row(i).sum
835 loss = win_loss_matrix.matrix.col(i).sum
837 yaml["players"][rating_group][p] =
838 { 'name' => p.split("+")[0],
839 'rating_group' => rating_group,
840 'rate' => rating.rate[i],
841 'last_modified' => $players_time[p].dup,
849 non_rated_group = 999 # large enough
850 yaml["players"][non_rated_group] = {}
851 $players.each_key do |id|
852 # skip players who have already been rated
854 (0..rating_group).each do |i|
855 found = true if yaml["players"][i][id]
860 v = GSL::Vector[0, 0]
861 $players[id].each_value {|value| v += value}
862 next if v[0] < 1 && v[1] < 1
864 yaml["players"][non_rated_group][id] =
865 { 'name' => id.split("+")[0],
866 'rating_group' => non_rated_group,
868 'last_modified' => $players_time[id].dup,
872 unless validate(yaml)
873 $stderr.puts "Aborted. It did not result in valid ratings."
874 $stderr.puts yaml.to_yaml if $DEBUG
884 # vim: ts=2 sw=2 sts=0