4 # Author:: Daigo Moriwaki
5 # Homepage:: http://sourceforge.jp/projects/shogi-server/
8 # Copyright (C) 2006-2009 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 genrated by the
39 # mk_game_results command.
40 # In the second style above, the file content can be read from the stdin.
43 # a base time point for this calicuration (default now). Ex. '2009-10-31'
46 # n [days] (default 60)
48 # --half-life-ignore::
49 # m [days] (default 7)
50 # after m days, the half-life effect works
53 # m [days] (default 365*2)
54 # old results will be ignored
56 # --fixed-rate-player::
57 # player whose rate is fixed at the rate
63 # skip draw games. [default: draw games are counted in as 0.5 win and 0.5
71 # Sample Command lines that isntall prerequires will work on Debian.
75 # $ sudo aptitude install ruby1.8
79 # $ sudo aptitude install rubygems
81 # * Ruby bindings for the GNU Scientific Library (GSL[http://rb-gsl.rubyforge.org/])
83 # $ sudo aptitude install libgsl-ruby1.8
85 # * RGL: {Ruby Graph Library}[http://rubyforge.org/projects/rgl/]
87 # $ sudo gem install rgl
91 # $ ./mk_rate game_results.txt > players.yaml
93 # $ ./mk_game_results . | ./mk_rate > players.yaml
95 # If you do not want the file to be update in case of errors,
97 # $ ./mk_rate game_results.txt && ./mk_rate game_results.txt > players.yaml
99 # == How players are rated
101 # The conditions that games and players are rated as following:
103 # * Rated games, which were played by both rated players.
104 # * Rated players, who logged in the server with a name followed by a trip: "name,trip".
105 # * (Rated) players, who played more than $GAMES_LIMIT [15] (rated) games.
113 require 'rgl/adjacency'
114 require 'rgl/connected_components'
116 #################################################
120 # Count out players who play less games than $GAMES_LIMIT
121 $GAMES_LIMIT = $DEBUG ? 0 : 15
128 # Holds the last time when a player gamed
129 $players_time = Hash.new { Time.at(0) }
132 #################################################
133 # Keeps the value of the lowest key
141 if @lowest.empty? || key < @lowest[0]
142 @lowest = [key, value]
155 #################################################
156 # Calculates rates of every player from a Win Loss GSL::Matrix
161 # The model of the win possibility is 1/(1 + 10^(-d/400)).
162 # The equation in this class is 1/(1 + e^(-Kd)).
163 # So, K should be calculated like this.
164 K = Math.log(10.0) / 400.0
166 # Convergence limit to stop Newton method.
168 # Stop Newton method after this iterations.
171 # Average rate among the players
180 # Calcurates the average of the vector.
182 def Rating.average(vector, mean=0.0)
183 sum = Array(vector).inject(0.0) {|sum, n| sum + n}
184 vector -= GSL::Vector[*Array.new(vector.size, sum/vector.size - mean)]
191 def initialize(win_loss_matrix)
195 when GSL::Matrix, GSL::Matrix::Int
204 attr_reader :rate, :n
208 (0...@size).collect {|k| yield k}
213 (0...@size).each {|k| yield k}
217 # The possibility that the player k will beet the player i.
220 1.0/(1.0 + exp(@rate[i]-@rate[k]))
224 # Most possible equation
231 sum += @n[k,i] * win_rate(i,k) - @n[i,k] * win_rate(k,i)
238 # / f0/R0 f0/R1 f0/R2 ... \
239 # dfk/dRj = | f1/R0 f1/R1 f1/R2 ... |
240 # \ f2/R0 f2/R1 f2/R2 ... /
246 sum += win_rate(i,k) * win_rate(k,i) * (@n[k,i] + @n[i,k])
250 sum = 2.0 * win_rate(j,k) * win_rate(k,j) * (@n[k,j] + @n[j,k])
256 # Jacobi matrix of the func().
262 (0...@size).collect do |k|
263 (0...@size).collect do |j|
271 # The initial value of the rate, which is of very importance for Newton
272 # method. This is based on my huristics; the higher the win probablity of
273 # a player is, the greater points he takes.
278 v = GSL::Vector[0, 0]
281 v += GSL::Vector[@n[k,i], @n[i,k]]
283 v.nrm2 < 1 ? 0 : v[0] / (v[0] + v[1])
285 rank = possibility.sort_index
286 @rate = player_vector do |k|
287 K*500 * (rank[k]+1) / @size
293 # Resets @rate as the higher the current win probablity of a player is,
294 # the greater points he takes.
297 @rate = @record.get || @rate
298 rank = @rate.sort_index
299 @rate = player_vector do |k|
300 K*@count*1.5 * (rank[k]+1) / @size
305 # mu is the deaccelrating parameter in Deaccelerated Newton method
306 def deaccelrate(mu, old_rate, a, old_f_nrm2)
307 @rate = old_rate - a * mu
308 if func_vector.nrm2 < (1 - mu / 4.0 ) * old_f_nrm2 then
312 @record.set(func_vector.nrm2, @rate)
316 $stderr.puts "mu: %f " % [mu] if $DEBUG
317 deaccelrate(mu*0.5, old_rate, a, old_f_nrm2)
321 # Main process to calculate ratings.
324 # Counter to stop the process.
325 # Calulation in Newton method may fall in an infinite loop
330 # Solve the equation:
332 # @rate_(n+1) = @rate_(n) - a
334 # f.nrm2 should approach to zero.
338 # $stderr.puts "j: %s" % [j.inspect] if $DEBUG
339 $stderr.puts "f: %s -> %f" % [f.to_a.inspect, f.nrm2] if $DEBUG
341 # GSL::Linalg::LU.solve or GSL::Linalg::HH.solve would be available instead.
342 #a = GSL::Linalg::HH.solve(j, f)
343 a, = GSL::MultiFit::linear(j, f)
344 a = self.class.average(a)
345 # $stderr.puts "a: %s -> %f" % [a.to_a.inspect, a.nrm2] if $DEBUG
347 # Deaccelerated Newton method
348 # GSL::Vector object should be immutable.
351 old_f_nrm2 = old_f.nrm2
352 deaccelrate(1.0, old_rate, a, old_f_nrm2)
353 #@rate -= a # Instead, do not deaccelerate
354 @record.set(func_vector.nrm2, @rate)
356 $stderr.printf "|error| : %5.2e\n", a.nrm2 if $DEBUG
359 if @count > COUNT_MAX
360 $stderr.puts "Values seem to oscillate. Stopped the process."
361 $stderr.puts "f: %s -> %f" % [func_vector.to_a.inspect, func_vector.nrm2]
365 end while (a.nrm2 > ERROR_LIMIT * @rate.nrm2)
368 $stderr.puts "resolved f: %s -> %f" %
369 [func_vector.to_a.inspect, func_vector.nrm2] if $DEBUG
370 $stderr.puts "Count: %d" % [@count] if $DEBUG
378 # Make the values of @rate finite.
381 @rate = @rate.collect do |a|
383 a.infinite? * AVERAGE_RATE * 100
391 # Flatten the values of @rate.
393 def average!(mean=0.0)
394 @rate = self.class.average(@rate, mean)
400 def translate!(value)
405 # Make the values of @rate integer.
408 @rate = @rate.collect do |a|
414 a.infinite? * AVERAGE_RATE * 100
420 #################################################
421 # Encapsulate a pair of keys and win loss matrix.
422 # - keys is an array of player IDs; [gps+123, foo+234, ...]
423 # - matrix holds games # where player i (row index) beats player j (column index).
424 # The row and column indexes match with the keys.
426 # This object should be immutable. If an internal state is being modified, a
427 # new object is always returned.
435 def self.mk_matrix(players)
436 keys = players.keys.sort
440 ((0...size).collect do |k|
442 p1_hash = players[p1]
443 ((0...size).collect do |j|
448 v = p1_hash[p2] || GSL::Vector[0,0]
453 return WinLossMatrix.new(keys, matrix)
456 def self.mk_win_loss_matrix(players)
457 obj = mk_matrix(players)
465 # an array of player IDs; [gps+123, foo+234, ...]
468 # matrix holds games # where player i (row index) beats player j (column index).
469 # The row and column indexes match with the keys.
472 def initialize(keys, matrix)
478 # Returns the size of the keys/matrix
489 # Removes players in a rows such as [1,3,5], and then returns a new
492 def delete_rows(rows)
493 rows = rows.sort.reverse
496 (0...size).each do |i|
497 next if rows.include?(i)
498 row = @matrix.row(i).clone
504 if copied_cols.size == 0
505 new_matrix = GSL::Matrix.new
507 new_matrix = GSL::Matrix[*copied_cols]
510 new_keys = @keys.clone
512 new_keys.delete_at(j)
515 return WinLossMatrix.new(new_keys, new_matrix)
519 # Removes players who do not pass a criteria to be rated, and returns a
523 $stderr.puts @keys.inspect if $DEBUG
524 $stderr.puts @matrix.inspect if $DEBUG
526 (0...size).each do |i|
531 if win < 1 || loss < 1 || win + loss < $GAMES_LIMIT
536 # The recursion ends if there is nothing to delete
537 return self if delete.empty?
539 new_obj = delete_rows(delete)
544 # Cuts self into connecting groups such as each player in a group has at least
545 # one game with other players in the group. Returns them as an array.
547 def connected_subsets
548 g = RGL::AdjacencyGraph.new
549 (0...size).each do |k|
550 (0...size).each do |i|
559 g.each_connected_component do |c|
562 new_keys << keys[v.to_s.to_i]
567 subsets = subsets.sort {|a,b| b.size <=> a.size}
569 result = subsets.collect do |keys|
572 ((0...keys.size).collect do |k|
573 p1 = @keys.index(keys[k])
574 ((0...keys.size).collect do |j|
578 p2 = @keys.index(keys[j])
583 WinLossMatrix.new(keys, matrix)
590 "size : #{@keys.size}" + "\n" +
591 @keys.inspect + "\n" +
598 #################################################
603 # After NHAFE_LIFE days value will get half.
604 # 0.693 is constant, where exp(0.693) ~ 0.5
606 if days < $options["half-life-ignore"]
609 Math::exp(-0.693/$options["half-life"]*(days-$options["half-life-ignore"]))
613 def _add_win_loss(winner, loser, time)
614 how_long_days = ($options["base-date"] - time)/(3600*24)
615 $players[winner] ||= Hash.new { GSL::Vector[0,0] }
616 $players[loser] ||= Hash.new { GSL::Vector[0,0] }
617 $players[winner][loser] += GSL::Vector[1.0*half_life(how_long_days),0]
618 $players[loser][winner] += GSL::Vector[0,1.0*half_life(how_long_days)]
621 def _add_draw(player1, player2, time)
622 how_long_days = ($options["base-date"] - time)/(3600*24)
623 $players[player1] ||= Hash.new { GSL::Vector[0,0] }
624 $players[player2] ||= Hash.new { GSL::Vector[0,0] }
625 $players[player1][player2] += GSL::Vector[0.5*half_life(how_long_days),0.5*half_life(how_long_days)]
626 $players[player2][player1] += GSL::Vector[0.5*half_life(how_long_days),0.5*half_life(how_long_days)]
629 def _add_time(player, time)
630 $players_time[player] = time if $players_time[player] < time
633 def add(black_mark, black_name, white_name, white_mark, time)
634 if black_mark == WIN_MARK && white_mark == LOSS_MARK
635 _add_win_loss(black_name, white_name, time)
636 elsif black_mark == LOSS_MARK && white_mark == WIN_MARK
637 _add_win_loss(white_name, black_name, time)
638 elsif black_mark == DRAW_MARK && white_mark == DRAW_MARK
639 if $options["skip-draw-games"]
642 _add_draw(black_name, white_name, time)
645 raise "Never reached!"
647 _add_time(black_name, time)
648 _add_time(white_name, time)
652 if /@NORATE\+/ =~ id # the player having @NORATE in the name should not be rated
655 id.gsub(/@.*?\+/,"+")
658 # Parse a game result line
661 time, state, black_mark, black_id, white_id, white_mark, file = line.split("\t")
662 unless time && state && black_mark && black_id &&
663 white_id && white_mark && file
664 $stderr.puts "Failed to parse the line : #{line}"
668 return if state == "abnormal"
669 time = Time.parse(time)
670 return if $options["base-date"] < time
671 how_long_days = ($options["base-date"] - time)/(3600*24)
672 if (how_long_days > $options["ignore"])
676 black_id = identify_id(black_id)
677 white_id = identify_id(white_id)
679 if black_id && white_id && (black_id != white_id) &&
680 black_mark && white_mark
681 add(black_mark, black_id, white_id, white_mark, time)
686 yaml["players"].each do |group_key, group|
687 group.each do |player_key, player|
688 rate = player['rate']
690 if rate > 10000 || rate < -10000
700 USAGE: #{$0} [options] GAME_RESULTS_FILE [...]
704 a path to a file listing results of games, which is genrated by the
705 mk_game_results command.
706 In the second style above, the file content can be read from the stdin.
709 --base-date a base time point for this calicuration (default now). Ex. '2009-10-31'
710 --half-life n [days] (default 60)
711 --half-life-ignore m [days] (default 7)
712 after m days, half-life effect works
713 --ignore n [days] (default 730 [=365*2]).
714 Results older than n days from the 'base-date' are ignored.
715 --fixed-rate-player player whose rate is fixed at the rate
717 --skip-draw-games skip draw games. [default: draw games are counted in
718 as 0.5 win and 0.5 lost]
719 --help show this message
725 parser = GetoptLong.new(
726 ["--base-date", GetoptLong::REQUIRED_ARGUMENT],
727 ["--half-life", GetoptLong::REQUIRED_ARGUMENT],
728 ["--half-life-ignore", GetoptLong::REQUIRED_ARGUMENT],
729 ["--help", "-h", GetoptLong::NO_ARGUMENT],
730 ["--ignore", GetoptLong::REQUIRED_ARGUMENT],
731 ["--fixed-rate-player", GetoptLong::REQUIRED_ARGUMENT],
732 ["--fixed-rate", GetoptLong::REQUIRED_ARGUMENT],
733 ["--skip-draw-games", GetoptLong::NO_ARGUMENT])
736 parser.each_option do |name, arg|
738 $options[name] = arg.dup
740 if ( $options["fixed-rate-player"] && !$options["fixed-rate"]) ||
741 (!$options["fixed-rate-player"] && $options["fixed-rate"]) ||
742 ( $options["fixed-rate-player"] && $options["fixed-rate"].to_i <= 0)
748 raise parser.error_message
754 if $options["base-date"]
755 $options["base-date"] = Time::parse $options["base-date"]
757 $options["base-date"] = Time.now
759 $options["half-life"] ||= 60
760 $options["half-life"] = $options["half-life"].to_i
761 $options["half-life-ignore"] ||= 7
762 $options["half-life-ignore"] = $options["half-life-ignore"].to_i
763 $options["ignore"] ||= 365*2
764 $options["ignore"] = $options["ignore"].to_i
765 $options["fixed-rate"] = $options["fixed-rate"].to_i if $options["fixed-rate"]
768 while line = $stdin.gets do
772 while file = ARGV.shift do
773 File.open(file) do |f|
774 f.each_line do |line|
785 obj = WinLossMatrix::mk_win_loss_matrix($players)
786 obj.connected_subsets.each do |win_loss_matrix|
787 yaml["players"][rating_group] = {}
789 rating = Rating.new(win_loss_matrix.matrix)
791 rating.average!(Rating::AVERAGE_RATE)
794 if $options["fixed-rate-player"]
795 # first, try exact match
796 index = win_loss_matrix.keys.index($options["fixed-rate-player"])
797 # second, try regular match
799 win_loss_matrix.keys.each_with_index do |p, i|
800 if %r!#{$options["fixed-rate-player"]}! =~ p
806 the_rate = rating.rate[index]
807 rating.translate!($options["fixed-rate"] - the_rate)
811 win_loss_matrix.keys.each_with_index do |p, i| # player_id, index#
812 win = win_loss_matrix.matrix.row(i).sum
813 loss = win_loss_matrix.matrix.col(i).sum
815 yaml["players"][rating_group][p] =
816 { 'name' => p.split("+")[0],
817 'rating_group' => rating_group,
818 'rate' => rating.rate[i],
819 'last_modified' => $players_time[p].dup,
827 non_rated_group = 999 # large enough
828 yaml["players"][non_rated_group] = {}
829 $players.each_key do |id|
830 # skip players who have already been rated
832 (0..rating_group).each do |i|
833 found = true if yaml["players"][i][id]
838 v = GSL::Vector[0, 0]
839 $players[id].each_value {|value| v += value}
840 next if v[0] < 1 && v[1] < 1
842 yaml["players"][non_rated_group][id] =
843 { 'name' => id.split("+")[0],
844 'rating_group' => non_rated_group,
846 'last_modified' => $players_time[id].dup,
850 unless validate(yaml)
851 $stderr.puts "Aborted. It did not result in valid ratings."
852 $stderr.puts yaml.to_yaml if $DEBUG
862 # vim: ts=2 sw=2 sts=0