Changeset 11747 for branches/HeuristicLab.Problems.GrammaticalOptimization/Main/Program.cs

Timestamp:

01/12/15 21:23:01 (10 years ago)

Author:

gkronber

Message:

#2283: implemented test problems for MCTS

File:

• branches/HeuristicLab.Problems.GrammaticalOptimization/Main/Program.cs (modified) (5 diffs)

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

@@ -88 +88 @@
           Tuple.Create((IProblem)new SymbolicRegressionPoly10Problem(), 23), 
         })
-        foreach (var randomTries in new int[] { 1, 10, /* 5, 100 /*, 500, 1000 */}) {
+        foreach (var randomTries in new int[] { 0, 1, 10, /* 5, 100 /*, 500, 1000 */}) {
           foreach (var policy in policies) {
             var myRandomTries = randomTries;
@@ -137 +137 @@
 
     private static void RunDemo() {
+      // TODO: unify MCTS, TD and ContextMCTS Solvers (stateInfos)
       // TODO: test with eps-greedy using max instead of average as value (seems to work well for symb-reg! explore further!)
       // TODO: separate value function from policy
@@ -165 +166 @@
       var random = new Random();
 
-      var problem = new SymbolicRegressionPoly10Problem();   // good results e.g. 10 randomtries and EpsGreedyPolicy(0.2, (aInfo)=>aInfo.MaxReward)
+      var phraseLen = 1;
+      var sentenceLen = 25;
+      var numPhrases = sentenceLen / phraseLen;
+      var problem = new RoyalPhraseSequenceProblem(random, 10, numPhrases, phraseLen: 1, k: 1, correctReward: 1, incorrectReward: 0);
+
+      //var problem = new SymbolicRegressionPoly10Problem();   // good results e.g. 10 randomtries and EpsGreedyPolicy(0.2, (aInfo)=>aInfo.MaxReward)
       // Ant
       // good results e.g. with       var alg = new MctsSampler(problem, 17, random, 1, (rand, numActions) => new ThresholdAscentPolicy(numActions, 500, 0.01));
@@ -175 +181 @@
       //var problem = new RoyalPairProblem();
       //var problem = new EvenParityProblem();
-      //var alg = new MctsSampler(problem, 23, random, 0, new GaussianThompsonSamplingPolicy(true));
-      var alg = new MctsContextualSampler(problem, 23, random, 0);
-      //var alg = new TemporalDifferenceTreeSearchSampler(problem, 17, random, 10, new EpsGreedyPolicy(0.1));
-      //var alg = new ExhaustiveBreadthFirstSearch(problem, 17);
+      // symbreg length = 11 q = 0.824522210419616
+      var alg = new MctsSampler(problem, sentenceLen, random, 0, new BoltzmannExplorationPolicy(200));
+      //var alg = new MctsQLearningSampler(problem, sentenceLen, random, 0, null);
+      //var alg = new MctsQLearningSampler(problem, 30, random, 0, new EpsGreedyPolicy(0.2));
+      //var alg = new MctsContextualSampler(problem, 23, random, 0); // must visit each canonical solution only once
+      //var alg = new TemporalDifferenceTreeSearchSampler(problem, 30, random, 1);
+      //var alg = new ExhaustiveBreadthFirstSearch(problem, 7);
       //var alg = new AlternativesContextSampler(problem, random, 17, 4, (rand, numActions) => new RandomPolicy(rand, numActions));
       //var alg = new ExhaustiveDepthFirstSearch(problem, 17);
       // var alg = new AlternativesSampler(problem, 17);
       // var alg = new RandomSearch(problem, random, 17);
-      // var alg = new ExhaustiveRandomFirstSearch(problem, random, 17);
+      //var alg = new ExhaustiveRandomFirstSearch(problem, random, 17);
 
       alg.FoundNewBestSolution += (sentence, quality) => {
@@ -199 +208 @@
           alg.PrintStats();
         }
+        //Console.WriteLine(sentence);
 
         if (iterations % 10000 == 0) {
-          //Console.WriteLine("{0,10} {1,10:F5} {2,10:F5} {3}", iterations, bestQuality, quality, sentence);
-          //Console.WriteLine("{0,4} {1,7} {2}", alg.treeDepth, alg.treeSize, globalStatistics);
           //Console.WriteLine("{0,4} {1,7} {2}", alg.treeDepth, alg.treeSize, globalStatistics);
         }