Changeset 11832 for branches/HeuristicLab.Problems.GrammaticalOptimization/Main/Program.cs

Timestamp:

01/27/15 16:34:34 (10 years ago)

Author:

gkronber

Message:

linear value function approximation and good results for poly-10 benchmark

File:

• branches/HeuristicLab.Problems.GrammaticalOptimization/Main/Program.cs (modified) (12 diffs)

branches/HeuristicLab.Problems.GrammaticalOptimization/Main/Program.cs

@@ -24 +24 @@
       CultureInfo.DefaultThreadCurrentCulture = CultureInfo.InvariantCulture;
 
-      RunDemo();
-      //RunGridTest();
+      //RunDemo();
+      RunGridTest();
     }
 
     private static void RunGridTest() {
-      int maxIterations = 50000; // for poly-10 with 50000 evaluations no successful try with hl yet
+      int maxIterations = 70000; // for poly-10 with 50000 evaluations no successful try with hl yet
       //var globalRandom = new Random(31415);
       var localRandSeed = 31415;
-      var reps = 10;
+      var reps = 30;
 
       var policyFactories = new Func<IBanditPolicy>[]
@@ -109 +109 @@
       {
         //(rand) => Tuple.Create((IProblem)new SantaFeAntProblem(), 17),
-        (rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:0, correctReward:1, decoyReward:0, phrasesAsSets:false ), 15),
-        (rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:0, correctReward:1, decoyReward:0, phrasesAsSets:true ), 15),
-        (rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:200, correctReward:1, decoyReward:0.5, phrasesAsSets:false), 15),
-        (rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:200, correctReward:1, decoyReward:0.5, phrasesAsSets:true), 15),
-        //(rand) => Tuple.Create((IProblem)new SymbolicRegressionPoly10Problem(), 23)
+        //(rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:0, correctReward:1, decoyReward:0, phrasesAsSets:false ), 15),
+        //(rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:0, correctReward:1, decoyReward:0, phrasesAsSets:true ), 15),
+        //(rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:200, correctReward:1, decoyReward:0.5, phrasesAsSets:false), 15),
+        //(rand) => Tuple.Create((IProblem)new FindPhrasesProblem(rand, 10, numPhrases:5, phraseLen:3, numOptimalPhrases:5, numDecoyPhrases:200, correctReward:1, decoyReward:0.5, phrasesAsSets:true), 15),
+        (rand) => Tuple.Create((IProblem)new SymbolicRegressionPoly10Problem(), 23)
       };
 
       foreach (var instanceFactory in instanceFactories) {
-        foreach (var useCanonical in new bool[] { true /*, false */ }) {
-          foreach (var randomTries in new int[] { 0, /* 1, 10, /* 5, 100 /*, 500, 1000 */}) {
+        foreach (var useCanonical in new bool[] { true /*, false */}) {
+          foreach (var randomTries in new int[] { 0 /*, 1, 10 /*, /* 5, 100 /*, 500, 1000 */}) {
             foreach (var policyFactory in policyFactories) {
               var myRandomTries = randomTries;
@@ -142 +142 @@
                 var problem = instance.Item1;
                 var maxLen = instance.Item2;
-                var alg = new SequentialSearch(problem, maxLen, myLocalRand, myRandomTries,
-                  new GenericGrammarPolicy(problem, policyFactory(), useCanonical));
+                //var alg = new SequentialSearch(problem, maxLen, myLocalRand, myRandomTries,
+                //  new GenericGrammarPolicy(problem, policyFactory(), useCanonical));
+                var alg = new SequentialSearch(problem, maxLen, myLocalRand,
+                  myRandomTries,
+                  new GenericFunctionApproximationGrammarPolicy(problem,
+                    useCanonical));
                 //var alg = new ExhaustiveBreadthFirstSearch(problem, 25);
                 //var alg = new AlternativesContextSampler(problem, 25);
@@ -150 +154 @@
                   iterations++;
                   globalStatistics.AddSentence(sentence, quality);
-                  if (iterations % 10000 == 0) {
-                    Console.WriteLine("{0,3} {1,5} \"{2,25}\" {3} {4}", i, myRandomTries, policyFactory(), useCanonical, globalStatistics);
+                  if (iterations % 1000 == 0) {
+                    Console.WriteLine("{0,3} {1,5} \"{2,25}\" {3} {4} {5}", i, myRandomTries, policyFactory(), useCanonical, problem.ToString(), globalStatistics);
                   }
                 };
@@ -190 +194 @@
 
 
-      int maxIterations = 100000;
+      int maxIterations = 1000000;
       int iterations = 0;
       var sw = new Stopwatch();
@@ -199 +203 @@
 
       //var problem = new RoyalSequenceProblem(random, 10, 30, 2, 1, 0);
+      // var phraseLen = 3;
+      // var numPhrases = 5;
+      // var problem = new RoyalPhraseSequenceProblem(random, 10, numPhrases, phraseLen: phraseLen, numCorrectPhrases: 1, correctReward: 1, incorrectReward: 0.0, phrasesAsSets: false);
+
       //var phraseLen = 3;
       //var numPhrases = 5;
-      //var problem = new RoyalPhraseSequenceProblem(random, 15, numPhrases, phraseLen: phraseLen, numCorrectPhrases: 1, correctReward: 1, incorrectReward: 0.0, phrasesAsSets: true);
-
-      // var phraseLen = 3;
-      // var numPhrases = 5;
-      // var problem = new FindPhrasesProblem(random, 10, numPhrases, phraseLen, numOptimalPhrases: numPhrases, numDecoyPhrases: 200, correctReward: 1.0, decoyReward: 0.5, phrasesAsSets: true);
+      //var problem = new FindPhrasesProblem(random, 10, numPhrases, phraseLen, numOptimalPhrases: numPhrases, numDecoyPhrases: 0, correctReward: 1.0, decoyReward: 0, phrasesAsSets: false);
 
       // good results for symb-reg
@@ -213 +217 @@
       // - GenericThompsonSamplingPolicy("")
       // - UCTPolicy(0.10) (5 of 5 runs, 35000 iters avg.), 10 successful runs of 10 with rand-tries 0, bei 40000 iters 9 / 10, bei 30000 1 / 10
+      // 2015 01 22: symb-reg: grid test on find-phrases problem showed good results for UCB1TunedPolicy and SequentialSearch with canonical states
+      // - symb-reg: consistent results with UCB1Tuned. finds optimal solution in ~50k iters (new GenericGrammarPolicy(problem, new UCB1TunedPolicy(), true));
+      // 2015 01 23: grid test with canonical states:
+      // - UCTPolicy(0.10) und UCBNormalPolicy 10/10 optimale Lösungen bei max. 50k iters, etwas schlechter: generic-thompson with variable sigma und bolzmannexploration (100)
+
 
       // good results for artificial ant:
@@ -219 +228 @@
       // - GaussianModelWithUnknownVariance (and Q= 0.99-quantil) also works well for Ant
       // 2015 01 19: grid test with canonical states (non-canonical slightly worse)
-      // - Threshold Ascent (best 100, 0.01; all variants relatively good)
-      // - Policies where the variance has a large weight compared to the mean? (Gaussian(compatible), Gaussian with fixed variance, UCT with large c, alle TA)
-
-      //var problem = new SymbolicRegressionPoly10Problem();
-
-      var problem = new SantaFeAntProblem();
-      //var problem = new SymbolicRegressionProblem("Tower"); 
+      // - ant: Threshold Ascent (best 100, 0.01; all variants relatively good)
+      // - ant: Policies where the variance has a large weight compared to the mean? (Gaussian(compatible), Gaussian with fixed variance, UCT with large c, alle TA)
+      // - ant: UCB1Tuned with canonical states also works very well for the artificial ant! constistent solutions in less than 10k iters      
+
+      var problem = new SymbolicRegressionPoly10Problem();
+      //var problem = new SantaFeAntProblem();
+      //var problem = new SymbolicRegressionProblem(random, "Tower"); 
       //var problem = new PalindromeProblem();
       //var problem = new HardPalindromeProblem();
@@ -234 +243 @@
       //var alg = new MctsSampler(problem, 23, random, 0, new EpsGreedyPolicy(0.1));
       //var alg = new SequentialSearch(problem, 23, random, 0,
-      //  new HeuristicLab.Algorithms.Bandits.GrammarPolicies.GenericGrammarPolicy(problem, new ModifiedUCTPolicy(0.1), true));
-      var alg = new SequentialSearch(problem, 17, random, 0,
-        new HeuristicLab.Algorithms.Bandits.GrammarPolicies.GenericTDPolicy(problem, true));
+      //  new HeuristicLab.Algorithms.Bandits.GrammarPolicies.QLearningGrammarPolicy(problem, new BoltzmannExplorationPolicy(10),
+      //    1, 1, true));
+      //var alg = new SequentialSearch(problem, 23, random, 0,
+      //  new HeuristicLab.Algorithms.Bandits.GrammarPolicies.GenericContextualGrammarPolicy(problem, new GenericThompsonSamplingPolicy(new GaussianModel(0.5, 10, 1, 1)), true));
+      var alg = new SequentialSearch(problem, 23, random, 0,
+        new HeuristicLab.Algorithms.Bandits.GrammarPolicies.GenericFunctionApproximationGrammarPolicy(problem, true));
       //var alg = new MctsQLearningSampler(problem, sentenceLen, random, 0, null);
       //var alg = new MctsQLearningSampler(problem, 30, random, 0, new EpsGreedyPolicy(0.2));
@@ -249 +261 @@
 
       alg.FoundNewBestSolution += (sentence, quality) => {
-        //Console.WriteLine("{0,4} {1,7} {2}", alg.treeDepth, alg.treeSize, globalStatistics);
+        //Console.WriteLine("{0}", globalStatistics);
         //Console.ReadLine();
       };
@@ -255 +267 @@
         iterations++;
         globalStatistics.AddSentence(sentence, quality);
+
         if (iterations % 1000 == 0) {
           if (iterations % 10000 == 0) Console.Clear();
@@ -260 +273 @@
           alg.PrintStats();
         }
+
         //Console.WriteLine(sentence);
 
-        if (iterations % 10000 == 0) {
-          //Console.WriteLine("{0,4} {1,7} {2}", alg.treeDepth, alg.treeSize, globalStatistics);
-        }
+        //if (iterations % 10000 == 0) {
+        //  Console.WriteLine("{0}", globalStatistics);
+        //}
       };