source: branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Problems.GrammaticalOptimization/FindPhrasesProblem.cs @ 11770

using System;
using System.Collections.Generic;
using System.Data.Odbc;
using System.Diagnostics;
using System.Linq;
using System.Text;
using System.Text.RegularExpressions;
using HeuristicLab.Common;

namespace HeuristicLab.Problems.GrammaticalOptimization {
  // must find a set of phrases where the ordering of phrases is irrelevant
  // Parameters
  // - size of the alphabet
  // - phrase length
  // - number of phrases in the sequence
  // - number of optimal phrases
  // - reward for optimal phrases
  // - number of decoy (sub-optimal) phrases
  // - reward for decoy phrases (must be smaller than reward for optimal phrases)
  // - phrasesAsSets: a switch to determine wether symbols in a phrase can be shuffled (sets) or if the ordering is relevant (non-sets)

  // this problem should be similar to symbolic regression and should be easier for approaches using a state esimation value and the canoncial state 
  // when phrases are symbol sets instead of sequences then value-estimation routines should be better (TD)
  public class FindPhrasesProblem : IProblem {

    private readonly IGrammar grammar;
    private readonly int numPhrases;
    private readonly int phraseLen;
    private readonly int numOptimalPhrases;
    private readonly int numDecoyPhrases;
    private readonly double correctReward;
    private readonly double decoyReward;
    private readonly bool phrasesAsSets;
    private readonly SortedSet<string> optimalPhrases;
    private readonly SortedSet<string> decoyPhrases;

    public FindPhrasesProblem(Random rand, int alphabetSize, int numPhrases, int phraseLen, int numOptimalPhrases, int numDecoyPhrases = 1,
      double correctReward = 1.0, double decoyReward = 0.0, bool phrasesAsSets = false) {
      if (alphabetSize <= 0 || alphabetSize > 26) throw new ArgumentException();
      if (numPhrases <= 0) throw new ArgumentException();
      if (phraseLen < 1) throw new ArgumentException();
      if (numOptimalPhrases < numPhrases) throw new ArgumentException();
      if (numDecoyPhrases < 0) throw new ArgumentException();
      if (correctReward <= decoyReward) throw new ArgumentException();

      this.numPhrases = numPhrases;
      this.phraseLen = phraseLen;
      this.correctReward = correctReward;
      this.decoyReward = decoyReward;
      this.phrasesAsSets = phrasesAsSets;

      // create grammar
      var sentenceSymbol = 'S';
      var terminalSymbols = Enumerable.Range(0, alphabetSize).Select(off => (char)((byte)'a' + off)).ToArray();
      var nonTerminalSymbols = new char[] { 'S' };
      var rules = terminalSymbols.Select(t => Tuple.Create('S', t.ToString()))
        .Concat(terminalSymbols.Select(t => Tuple.Create('S', t + "S")));

      this.grammar = new Grammar(sentenceSymbol, terminalSymbols, nonTerminalSymbols, rules);

      // generate optimal phrases
      optimalPhrases = new SortedSet<string>();
      while (optimalPhrases.Count < numOptimalPhrases) {
        string phrase = "";
        for (int l = 0; l < phraseLen; l++) {
          phrase += terminalSymbols.SelectRandom(rand);
        }
        phrase = CanonicalPhrase(phrase);

        // don't allow dups
        if (!optimalPhrases.Contains(phrase)) optimalPhrases.Add(phrase);
      }

      // generate decoy phrases
      decoyPhrases = new SortedSet<string>();
      while (decoyPhrases.Count < numDecoyPhrases) {
        string phrase = "";
        for (int l = 0; l < phraseLen; l++) {
          phrase += terminalSymbols.SelectRandom(rand);
        }
        phrase = CanonicalPhrase(phrase);

        // don't allow dups
        if (!optimalPhrases.Contains(phrase) && !decoyPhrases.Contains(phrase)) decoyPhrases.Add(phrase);
      }

      Debug.Assert(Evaluate(BestKnownSolution) / BestKnownQuality(phraseLen * numPhrases) == 1.0);
    }

    public double BestKnownQuality(int maxLen) {
      return Math.Min(maxLen / phraseLen, numPhrases) * correctReward; // integer division
    }

    public string BestKnownSolution {
      get { return string.Join("", optimalPhrases.Take(numPhrases)); }
    }

    public IGrammar Grammar {
      get { return grammar; }
    }

    public double Evaluate(string sentence) {
      // sentence must contain only terminal symbols, we are not checking if the sentence is syntactically valid here because it would be too slow!
      Debug.Assert(sentence.Any(c => grammar.IsTerminal(c)));


      // split the sentence in phrases
      // phrases must not overlap in the sentence, multiple occurences of a phrase are not counted
      // the order of phrases is not relevant
      var numPhrases = sentence.Length / phraseLen;
      var phrases = new SortedSet<string>();
      for (int phraseIdx = 0; phraseIdx < numPhrases; phraseIdx++) {
        var sentenceIdx = phraseIdx * phraseLen;
        var phrase = sentence.Substring(sentenceIdx, phraseLen);
        phrase = CanonicalPhrase(phrase);
        if (!phrases.Contains(phrase)) phrases.Add(phrase);
      }

      // add reward for each correct phrase that occurs in the sentence 
      // add reward for each decoy phrase that occurs in the sentence
      var reward = phrases.Intersect(optimalPhrases).Count() * correctReward
               + phrases.Intersect(decoyPhrases).Count() * decoyReward;



      return reward;
    }

    private string CanonicalPhrase(string phrase) {
      if (phrasesAsSets) return string.Join("", phrase.OrderBy(ch => (byte)ch));
      else return phrase;
    }

    public string CanonicalRepresentation(string terminalPhrase) {
      // as the ordering of phrases does not matter we can reorder the phrases
      // and remove duplicates
      var numPhrases = terminalPhrase.Length / phraseLen;
      var phrases = new SortedSet<string>();
      for (int phraseIdx = 0; phraseIdx < numPhrases; phraseIdx++) {
        var sentenceIdx = phraseIdx * phraseLen;
        var phrase = terminalPhrase.Substring(sentenceIdx, phraseLen);
        phrase = CanonicalPhrase(phrase);
        if (!phrases.Contains(phrase)) phrases.Add(phrase);
      }
      var remainder = terminalPhrase.Substring(numPhrases * phraseLen, terminalPhrase.Length - (numPhrases * phraseLen));
      remainder = CanonicalPhrase(remainder);
      if (!phrases.Contains(remainder)) phrases.Add(remainder);

      return string.Join("", phrases);
    }
  }
}