From: Ruby Quiz <james@...>
Date: 2007-07-12T06:34:24+09:00
Subject: [SUMMARY] Hangman (#130)

People wrote quite a bit of code for this quiz.  I'm guessing that's because it
can be challenging to see where how your code is doing and to refine that.

I'm going to show my own solution this time, not because it's any better than
the others, but because I can talk you through my thinking in building it.

The first thing in my code is a tiny library that handles dictionary loading for
the guessing script and the test script.  I extracted this common functionality
when I built my testing script:

	#!/usr/bin/env ruby -wKU
	
	WORDS_CASH_FILE = "words.cache"
	
	if File.exist? WORDS_CASH_FILE
	  WORDS = File.open(WORDS_CASH_FILE) { |file| Marshal.load(file) }
	else
	  WORDS = File.open( ARGV.find { |arg| arg =~ /\A[^-]/ } ||
	                     "/usr/share/dict/words" ) do |dict|
	    dict.inject(Hash.new) do |all, word|
	      all.update(word.delete("^A-Za-z").downcase => true)
	    end.keys.sort_by { |w| [w.length, w] }
	  end
	  File.open(WORDS_CASH_FILE, "w") { |file| Marshal.dump(WORDS, file) }
	end

There's nothing too fancy here.  Originally I just reread the dictionary every
time, but that slowed down testing too much and I found caching it with Marshal
helped.  I also started sorting the dictionary eventually to make it easier for
when my testing was, "I've tried everything beneath four letters and I'm in the
K's on those."

On to the guessing script.  Here's how that begins:

	#!/usr/bin/env ruby -wKU

	puts "One moment..."
	puts
	require "words"

	def frequency(words)
	  freq = Hash.new(0)
	  words.each do |word|
	    word.split("").each { |letter| freq[letter] += word.count(letter) }
	  end
	  freq
	end
	FREQ = frequency(WORDS).sort_by { |_, count| -count }.
	                        map { |letter, _| letter }

	choices = WORDS
	guesses = Array.new
	
	# ...

This code prints a warning that loading the dictionary might take a moment, then
requires the mini-library I just showed.  Once loaded, it defines a method for
counting letter frequencies and runs it on the dictionary.

After that it initializes some state variables for the guessing.  The script
decreases the word list as it works, so that is copied into a variable to
reflect that it will change.  We then setup a way to keep track of our guesses.

The next bit of code begins the main event loop:

	# ...
	
	loop do
	  puts guesses.empty?                                       ?
	       "Please enter a word pattern (_ _ _ _ for example):" :
	       "Please update your pattern according to my guess "  +
	       "(_ i _ _ for example):"
	  $stdout.flush
	  pattern = $stdin.gets.to_s.delete("^A-Za-z_")
	  
	  # ...

This code just asks the user for a pattern and retrieves what they give us.  The
call to flush() is to support the test script, which we will see in a bit.

	  # ...
	  
	  bad_guesses = guesses - pattern.delete("_").split("")
	  if bad_guesses.size > 5 and pattern.include? "_"
	    puts "I'm out of guesses.  You win."
	  elsif not pattern.include? "_"
	    puts "I guessed your word.  Pretty smart, huh?"
	  else
	    choices = choices.grep(
	                bad_guesses.empty?             ?
	                /\A#{pattern.tr("_", ".")}\Z/i :
	                /\A(?!.*[#{bad_guesses.join}])#{pattern.tr("_", ".")}\Z/i
	              )
	    guess = frequency(choices).
	              reject { |letter, _| guesses.include? letter }.
	              sort_by { |letter, count| [-count, FREQ.index(letter)] }.
	              first.first rescue nil

	    guesses << guess
	    puts "I guess the letter '#{guess}'."
	    puts
	    next
	  end
	  
	  # ...

This is the main body of this guessing script.  The first two branches just
check for win and loss conditions, but the else branch is where the logic is.

Before a selection is made, the dictionary is reduced by all words that couldn't
possibly match.  After that, a frequency count is made for all remaining
possibilities.  Those possible letters are then sorted by that count and how
common they are in this dictionary.  The letter that bubbles to the top is our
choice.

This selection process works decently on larger words.  It really starts to show
decent success rates at words four letters and longer.  Sadly, it fairs poorly
with smaller words.

Once chosen a guess is printed for the user and the loop cycles to the next
iteration.

When we fall through the above code we have reached an end condition.  The
following code sorts that out:

	  # ...
	  
	  puts
	  if ARGV.include? "--loop"
	    choices = WORDS
	    guesses = Array.new
	  else
	    break
	  end
	end

Here I watch for a special command-line switch that puts the program into a
continuous loop for testing.  This avoids reloading the dictionary and thus is
faster.

I landed on my word choice strategy mainly by trial and error.  That was made
possible by being able to run the algorithm over some words and watching how
accurately it could pick them.  I did that by writing a simple script that
pretend to be me.  It runs the game, feeds it words, and keeps score of it's
progress.  Here's the start of that code:

	#!/usr/bin/env ruby -wKU
	
	require "words"
	
	results = Hash.new(0)
	at_exit do
	  results[:total] = results[:right] + results[:wrong]
	  puts
	  puts   "Words:     #{results[:total]}"
	  puts   "Guessed:   #{results[:right]}"
	  puts   "Missed:    #{results[:wrong]}"
	  printf "Accuracy:  %.2f%%\n", results[:right] / results[:total].to_f * 100
	  puts
	end
	trap("INT") { exit }
	
	# ...

Here I load the dictionary using the same code the guesser relies on.  Then I
build a Hash to hold guess counts and setup result printing for when the program
exits.  This allows me to cancel test runs at anytime, but still see results
thus far.  That was nice for spot checking results.  Sometimes I could tell
right away that a change was better or worse.

Here's the actual testing code:

	# ...
	
	IO.popen( File.join(File.dirname(__FILE__), "hangman.rb --loop"),
	          "r+" ) do |hangman|
	  WORDS.each do |word|
	    pattern = word.tr("a-z", "_")
	    loop do
	      input = String.new
	      hangman.each do |line|
	        input << line
	        break if input =~ /^(?:I'm out|I guessed)|:\Z/
	      end
	
	      if input =~ /^I'm out/
	        puts "It missed '#{word}'."
	        results[:wrong] += 1
	        break
	      elsif input =~ /^I guessed/
	        puts "It guessed '#{word}'."
	        results[:right] += 1
	        break
	      elsif input =~ /^I guess the letter '(.)'/
	        guess = $1
	        word.split("").each_with_index do |letter, i|
	          pattern[i, 1] = letter if letter == guess
	        end
	      end
	
	      hangman.puts pattern
	    end
	  end
	end

This code just runs the guessing script just as a human would do.  It feeds the
script words and watches the responses to see what happens.  You can think of
this process as a poor man's expect.

Many of the other solutions achieved better results than my own, so be sure to
look through them.  Some were quite long but most had decent documentation about
their process.

My thanks to all who found the time to dig into this problem.

Tomorrow we have a classic computer science exercise and it's an easy one, so
warm up your impress-people-with-a-concise-solution skills...