source: branches/MCTS-SymbReg-2796/HeuristicLab.Algorithms.DataAnalysis/3.4/MctsSymbolicRegression/MctsSymbolicRegressionStatic.cs @ 15440

#region License Information
/* HeuristicLab
 * Copyright (C) 2002-2016 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
 * 
 * This file is part of HeuristicLab.
 *
 * HeuristicLab is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * HeuristicLab is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
 */
#endregion

using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using System.Text;
using HeuristicLab.Core;
using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
using HeuristicLab.Optimization;
using HeuristicLab.Problems.DataAnalysis;
using HeuristicLab.Problems.DataAnalysis.Symbolic;
using HeuristicLab.Problems.DataAnalysis.Symbolic.Regression;
using HeuristicLab.Random;

namespace HeuristicLab.Algorithms.DataAnalysis.MctsSymbolicRegression {
  public static class MctsSymbolicRegressionStatic {
    // OBJECTIVES:
    // 1) solve toy problems without numeric constants (to show that structure search is effective / efficient)
    //    - e.g. Keijzer, Nguyen ... where no numeric constants are involved
    //    - assumptions: 
    //      - we don't know the necessary operations or functions -> all available functions could be necessary
    //      - but we do not need to tune numeric constants -> no scaling of input variables x!
    // 2) Solve toy problems with numeric constants to make the algorithm invariant concerning variable scale.
    //    This is important for real world applications.
    //    - e.g. Korns or Vladislavleva problems where numeric constants are involved
    //    - assumptions:
    //      - any numeric constant is possible (a-priori we might assume that small abs. constants are more likely)
    //      - standardization of variables is possible (or might be necessary) as we adjust numeric parameters of the expression anyway
    //      - to simplify the problem we can restrict the set of functions e.g. we assume which functions are necessary for the problem instance
    //        -> several steps: (a) polyinomials, (b) rational polynomials, (c) exponential or logarithmic functions, rational functions with exponential and logarithmic parts
    // 3) efficiency and effectiveness for real-world problems
    //    - e.g. Tower problem
    //    - (1) and (2) combined, structure search must be effective in combination with numeric optimization of constants
    //    

    // TODO: The samples of x1*... or x2*... do not give any information about the relevance of the interaction term x1*x2 in general!
    //       --> E.g. if x1, x2 ~ N(0, 1) or U(-1, 1) this is trivial to show
    //       --> Therefore, looking at rollout statistics for arm selection is useless in the general case!
    //       --> It is necessary to rely on other features for the arm selection. 
    //       --> TODO: Which heuristics can we apply? 
    // TODO: Solve Poly-10
    // TODO: rename everything as this is not MCTS anymore
    // TODO: After state unification the recursive backpropagation of results takes a lot of time. How can this be improved?
    // ~~obsolete TODO: Why is the algorithm so slow for rather greedy policies (e.g. low C value in UCB)?
    // ~~obsolete TODO: check if we can use a quality measure with range [-1..1] in policies
    // TODO: unit tests for benchmark problems which contain log / exp / x^-1 but without numeric constants 
    // TODO: check if transformation of y is correct and works (Obj 2)
    // TODO: The algorithm is not invariant to location and scale of variables. 
    //       Include offset for variables as parameter (for Objective 2)
    // TODO: why does LM optimization converge so slowly with exp(x), log(x), and 1/x allowed (Obj 2)?
    // TODO: support e(-x) and possibly (1/-x) (Obj 1)
    // TODO: is it OK to initialize all constants to 1 (Obj 2)?
    // TODO: improve memory usage
    // TODO: support empty test partition
    // TODO: the algorithm should be invariant to linear transformations of the space (y = f(x') = f( Ax ) ) for invertible transformations A --> unit tests
    #region static API

    public interface IState {
      bool Done { get; }
      ISymbolicRegressionModel BestModel { get; }
      double BestSolutionTrainingQuality { get; }
      double BestSolutionTestQuality { get; }
      IEnumerable<ISymbolicRegressionSolution> ParetoBestModels { get; }
      int TotalRollouts { get; }
      int EffectiveRollouts { get; }
      int FuncEvaluations { get; }
      int GradEvaluations { get; } // number of gradient evaluations (* num parameters) to get a value representative of the effort comparable to the number of function evaluations
      // TODO other stats on LM optimizer might be interesting here
    }

    // created through factory method
    private class State : IState {
      private const int MaxParams = 100;

      // state variables used by MCTS
      internal readonly Automaton automaton;
      internal IRandom random { get; private set; }
      internal readonly Tree tree;
      internal readonly Func<byte[], int, double> evalFun;
      // MCTS might get stuck. Track statistics on the number of effective rollouts
      internal int totalRollouts;
      internal int effectiveRollouts;


      // state variables used only internally (for eval function)
      private readonly IRegressionProblemData problemData;
      private readonly double[][] x;
      private readonly double[] y;
      private readonly double[][] testX;
      private readonly double[] testY;
      private readonly double[] scalingFactor;
      private readonly double[] scalingOffset;
      private readonly double yStdDev; // for scaling parameters (e.g. stopping condition for LM)
      private readonly int constOptIterations;
      private readonly double lambda; // weight of penalty term for regularization
      private readonly double lowerEstimationLimit, upperEstimationLimit;
      private readonly bool collectParetoOptimalModels;
      private readonly List<ISymbolicRegressionSolution> paretoBestModels = new List<ISymbolicRegressionSolution>();
      private readonly List<double[]> paretoFront = new List<double[]>(); // matching the models

      private readonly ExpressionEvaluator evaluator, testEvaluator;

      internal readonly Dictionary<Tree, List<Tree>> children = new Dictionary<Tree, List<Tree>>();
      internal readonly Dictionary<Tree, List<Tree>> parents = new Dictionary<Tree, List<Tree>>();
      internal readonly Dictionary<ulong, Tree> nodes = new Dictionary<ulong, Tree>();

      // values for best solution
      private double bestR;
      private byte[] bestCode;
      private int bestNParams;
      private double[] bestConsts;

      // stats
      private int funcEvaluations;
      private int gradEvaluations;

      // buffers
      private readonly double[] ones; // vector of ones (as default params)
      private readonly double[] constsBuf;
      private readonly double[] predBuf, testPredBuf;
      private readonly double[][] gradBuf;

      public State(IRegressionProblemData problemData, uint randSeed, int maxVariables, bool scaleVariables,
        int constOptIterations, double lambda,
        bool collectParetoOptimalModels = false,
        double lowerEstimationLimit = double.MinValue, double upperEstimationLimit = double.MaxValue,
        bool allowProdOfVars = true,
        bool allowExp = true,
        bool allowLog = true,
        bool allowInv = true,
        bool allowMultipleTerms = false) {

        if (lambda < 0) throw new ArgumentException("Lambda must be larger or equal zero", "lambda");

        this.problemData = problemData;
        this.constOptIterations = constOptIterations;
        this.lambda = lambda;
        this.evalFun = this.Eval;
        this.lowerEstimationLimit = lowerEstimationLimit;
        this.upperEstimationLimit = upperEstimationLimit;
        this.collectParetoOptimalModels = collectParetoOptimalModels;

        random = new MersenneTwister(randSeed);

        // prepare data for evaluation
        double[][] x;
        double[] y;
        double[][] testX;
        double[] testY;
        double[] scalingFactor;
        double[] scalingOffset;
        // get training and test datasets (scale linearly based on training set if required)
        GenerateData(problemData, scaleVariables, problemData.TrainingIndices, out x, out y, out scalingFactor, out scalingOffset);
        GenerateData(problemData, problemData.TestIndices, scalingFactor, scalingOffset, out testX, out testY);
        this.x = x;
        this.y = y;
        this.yStdDev = HeuristicLab.Common.EnumerableStatisticExtensions.StandardDeviation(y);
        this.testX = testX;
        this.testY = testY;
        this.scalingFactor = scalingFactor;
        this.scalingOffset = scalingOffset;
        this.evaluator = new ExpressionEvaluator(y.Length, lowerEstimationLimit, upperEstimationLimit);
        // we need a separate evaluator because the vector length for the test dataset might differ 
        this.testEvaluator = new ExpressionEvaluator(testY.Length, lowerEstimationLimit, upperEstimationLimit);

        this.automaton = new Automaton(x, allowProdOfVars, allowExp, allowLog, allowInv, allowMultipleTerms, maxVariables);
        this.tree = new Tree() {
          state = automaton.CurrentState,
          expr = "",
          level = 0
        };

        // reset best solution
        this.bestR = 0;
        // code for default solution (constant model)
        this.bestCode = new byte[] { (byte)OpCodes.LoadConst0, (byte)OpCodes.Exit };
        this.bestNParams = 0;
        this.bestConsts = null;

        // init buffers
        this.ones = Enumerable.Repeat(1.0, MaxParams).ToArray();
        constsBuf = new double[MaxParams];
        this.predBuf = new double[y.Length];
        this.testPredBuf = new double[testY.Length];

        this.gradBuf = Enumerable.Range(0, MaxParams).Select(_ => new double[y.Length]).ToArray();
      }

      #region IState inferface
      public bool Done { get { return tree != null && tree.Done; } }

      public double BestSolutionTrainingQuality {
        get {
          evaluator.Exec(bestCode, x, bestConsts, predBuf);
          return Rho(y, predBuf);
        }
      }

      public double BestSolutionTestQuality {
        get {
          testEvaluator.Exec(bestCode, testX, bestConsts, testPredBuf);
          return Rho(testY, testPredBuf);
        }
      }

      // takes the code of the best solution and creates and equivalent symbolic regression model
      public ISymbolicRegressionModel BestModel {
        get {
          var treeGen = new SymbolicExpressionTreeGenerator(problemData.AllowedInputVariables.ToArray());
          var interpreter = new SymbolicDataAnalysisExpressionTreeLinearInterpreter();

          var t = new SymbolicExpressionTree(treeGen.Exec(bestCode, bestConsts, bestNParams, scalingFactor, scalingOffset));
          var model = new SymbolicRegressionModel(problemData.TargetVariable, t, interpreter, lowerEstimationLimit, upperEstimationLimit);
          model.Scale(problemData); // apply linear scaling
          return model;
        }
      }
      public IEnumerable<ISymbolicRegressionSolution> ParetoBestModels {
        get { return paretoBestModels; }
      }

      public int TotalRollouts { get { return totalRollouts; } }
      public int EffectiveRollouts { get { return effectiveRollouts; } }
      public int FuncEvaluations { get { return funcEvaluations; } }
      public int GradEvaluations { get { return gradEvaluations; } } // number of gradient evaluations (* num parameters) to get a value representative of the effort comparable to the number of function evaluations

      #endregion


#if DEBUG
      public string ExprStr(Automaton automaton) {
        byte[] code;
        int nParams;
        automaton.GetCode(out code, out nParams);
        var generator = new SymbolicExpressionTreeGenerator(problemData.AllowedInputVariables.ToArray());
        var @params = Enumerable.Repeat(1.0, nParams).ToArray();
        var root = generator.Exec(code, @params, nParams, null, null);
        var formatter = new InfixExpressionFormatter();
        return formatter.Format(new SymbolicExpressionTree(root));
      }
#endif

      private double Eval(byte[] code, int nParams) {
        double[] optConsts;
        double q;
        Eval(code, nParams, out q, out optConsts);

        // single objective best
        if (q > bestR) {
          bestR = q;
          bestNParams = nParams;
          this.bestCode = new byte[code.Length];
          this.bestConsts = new double[bestNParams];

          Array.Copy(code, bestCode, code.Length);
          Array.Copy(optConsts, bestConsts, bestNParams);
        }
        if (collectParetoOptimalModels) {
          // multi-objective best
          var complexity = // SymbolicDataAnalysisModelComplexityCalculator.CalculateComplexity() TODO: implement Kommenda's tree complexity directly in the evaluator
            Array.FindIndex(code, (opc) => opc == (byte)OpCodes.Exit);  // use length of expression as surrogate for complexity
          UpdateParetoFront(q, complexity, code, optConsts, nParams, scalingFactor, scalingOffset);
        }
        return q;
      }

      private void Eval(byte[] code, int nParams, out double rho, out double[] optConsts) {
        // we make a first pass to determine a valid starting configuration for all constants
        // constant c in log(c + f(x)) is adjusted to guarantee that x is positive (see expression evaluator)
        // scale and offset are set to optimal starting configuration
        // assumes scale is the first param and offset is the last param

        // reset constants
        Array.Copy(ones, constsBuf, nParams);
        evaluator.Exec(code, x, constsBuf, predBuf, adjustOffsetForLogAndExp: true);
        funcEvaluations++;

        if (nParams == 0 || constOptIterations < 0) {
          // if we don't need to optimize parameters then we are done
          // changing scale and offset does not influence r²
          rho = Rho(y, predBuf);
          optConsts = constsBuf;
        } else {
          // optimize constants using the starting point calculated above
          OptimizeConstsLm(code, constsBuf, nParams, 0.0, nIters: constOptIterations);

          evaluator.Exec(code, x, constsBuf, predBuf);
          funcEvaluations++;

          rho = Rho(y, predBuf);
          optConsts = constsBuf;
        }
      }



      #region helpers
      private static double Rho(IEnumerable<double> x, IEnumerable<double> y) {
        OnlineCalculatorError error;
        double r = OnlinePearsonsRCalculator.Calculate(x, y, out error);
        return error == OnlineCalculatorError.None ? r : 0.0;
      }


      private void OptimizeConstsLm(byte[] code, double[] consts, int nParams, double epsF = 0.0, int nIters = 100) {
        double[] optConsts = new double[nParams]; // allocate a smaller buffer for constants opt (TODO perf?)
        Array.Copy(consts, optConsts, nParams);

        // direct usage of LM is recommended in alglib manual for better performance than the lsfit interface (which uses lm internally).
        alglib.minlmstate state;
        alglib.minlmreport rep = null;
        alglib.minlmcreatevj(y.Length + 1, optConsts, out state); // +1 for penalty term
        // Using the change of the gradient as stopping criterion is recommended in alglib manual.
        // However, the most recent version of alglib (as of Oct 2017) only supports epsX as stopping criterion
        alglib.minlmsetcond(state, epsg: 1E-6 * yStdDev, epsf: epsF, epsx: 0.0, maxits: nIters);
        // alglib.minlmsetgradientcheck(state, 1E-5);
        alglib.minlmoptimize(state, Func, FuncAndJacobian, null, code);
        alglib.minlmresults(state, out optConsts, out rep);
        funcEvaluations += rep.nfunc;
        gradEvaluations += rep.njac * nParams;

        if (rep.terminationtype < 0) throw new ArgumentException("lm failed: termination type = " + rep.terminationtype);

        // only use optimized constants if successful
        if (rep.terminationtype >= 0) {
          Array.Copy(optConsts, consts, optConsts.Length);
        }
      }

      private void Func(double[] arg, double[] fi, object obj) {
        var code = (byte[])obj;
        int n = predBuf.Length;
        evaluator.Exec(code, x, arg, predBuf); // gradients are nParams x vLen
        for (int r = 0; r < n; r++) {
          var res = predBuf[r] - y[r];
          fi[r] = res;
        }

        var penaltyIdx = fi.Length - 1;
        fi[penaltyIdx] = 0.0;
        // calc length of parameter vector for regularization
        var aa = 0.0;
        for (int i = 0; i < arg.Length; i++) {
          aa += arg[i] * arg[i];
        }
        if (lambda > 0 && aa > 0) {
          // scale lambda using stdDev(y) to make the parameter independent of the scale of y
          // scale lambda using n to make parameter independent of the number of training points
          // take the root because LM squares the result
          fi[penaltyIdx] = Math.Sqrt(n * lambda / yStdDev * aa);
        }
      }

      private void FuncAndJacobian(double[] arg, double[] fi, double[,] jac, object obj) {
        int n = predBuf.Length;
        int nParams = arg.Length;
        var code = (byte[])obj;
        evaluator.ExecGradient(code, x, arg, predBuf, gradBuf); // gradients are nParams x vLen
        for (int r = 0; r < n; r++) {
          var res = predBuf[r] - y[r];
          fi[r] = res;

          for (int k = 0; k < nParams; k++) {
            jac[r, k] = gradBuf[k][r];
          }
        }
        // calc length of parameter vector for regularization
        double aa = 0.0;
        for (int i = 0; i < arg.Length; i++) {
          aa += arg[i] * arg[i];
        }

        var penaltyIdx = fi.Length - 1;
        if (lambda > 0 && aa > 0) {
          fi[penaltyIdx] = 0.0;
          // scale lambda using stdDev(y) to make the parameter independent of the scale of y
          // scale lambda using n to make parameter independent of the number of training points
          // take the root because alglib LM squares the result
          fi[penaltyIdx] = Math.Sqrt(n * lambda / yStdDev * aa);

          for (int i = 0; i < arg.Length; i++) {
            jac[penaltyIdx, i] = 0.5 / fi[penaltyIdx] * 2 * n * lambda / yStdDev * arg[i];
          }
        } else {
          fi[penaltyIdx] = 0.0;
          for (int i = 0; i < arg.Length; i++) {
            jac[penaltyIdx, i] = 0.0;
          }
        }
      }


      private void UpdateParetoFront(double q, int complexity, byte[] code, double[] param, int nParam,
        double[] scalingFactor, double[] scalingOffset) {
        double[] best = new double[2];
        double[] cur = new double[2] { q, complexity };
        bool[] max = new[] { true, false };
        var isNonDominated = true;
        foreach (var e in paretoFront) {
          var domRes = DominationCalculator<int>.Dominates(cur, e, max, true);
          if (domRes == DominationResult.IsDominated) {
            isNonDominated = false;
            break;
          }
        }
        if (isNonDominated) {
          paretoFront.Add(cur);

          // create model
          var treeGen = new SymbolicExpressionTreeGenerator(problemData.AllowedInputVariables.ToArray());
          var interpreter = new SymbolicDataAnalysisExpressionTreeLinearInterpreter();

          var t = new SymbolicExpressionTree(treeGen.Exec(code, param, nParam, scalingFactor, scalingOffset));
          var model = new SymbolicRegressionModel(problemData.TargetVariable, t, interpreter, lowerEstimationLimit, upperEstimationLimit);
          model.Scale(problemData); // apply linear scaling

          var sol = model.CreateRegressionSolution(this.problemData);
          sol.Name = string.Format("{0:N5} {1}", q, complexity);

          paretoBestModels.Add(sol);
        }
        for (int i = paretoFront.Count - 2; i >= 0; i--) {
          var @ref = paretoFront[i];
          var domRes = DominationCalculator<int>.Dominates(cur, @ref, max, true);
          if (domRes == DominationResult.Dominates) {
            paretoFront.RemoveAt(i);
            paretoBestModels.RemoveAt(i);
          }
        }
      }

      #endregion


    }


    /// <summary>
    /// Static method to initialize a state for the algorithm
    /// </summary>
    /// <param name="problemData">The problem data</param>
    /// <param name="randSeed">Random seed.</param>
    /// <param name="maxVariables">Maximum number of variable references that are allowed in the expression.</param>
    /// <param name="scaleVariables">Optionally scale input variables to the interval [0..1] (recommended)</param>
    /// <param name="constOptIterations">Maximum number of iterations for constants optimization (Levenberg-Marquardt)</param>
    /// <param name="lambda">Penalty factor for regularization (0..inf.), small penalty disabled regularization.</param>
    /// <param name="policy">Tree search policy (random, ucb, eps-greedy, ...)</param>
    /// <param name="collectParameterOptimalModels">Optionally collect all Pareto-optimal solutions having minimal length and error.</param>
    /// <param name="lowerEstimationLimit">Optionally limit the result of the expression to this lower value.</param>
    /// <param name="upperEstimationLimit">Optionally limit the result of the expression to this upper value.</param>
    /// <param name="allowProdOfVars">Allow products of expressions.</param>
    /// <param name="allowExp">Allow expressions with exponentials.</param>
    /// <param name="allowLog">Allow expressions with logarithms</param>
    /// <param name="allowInv">Allow expressions with 1/x</param>
    /// <param name="allowMultipleTerms">Allow expressions which are sums of multiple terms.</param>
    /// <returns></returns>

    public static IState CreateState(IRegressionProblemData problemData, uint randSeed, int maxVariables = 3,
      bool scaleVariables = true, int constOptIterations = -1, double lambda = 0.0,
      bool collectParameterOptimalModels = false,
      double lowerEstimationLimit = double.MinValue, double upperEstimationLimit = double.MaxValue,
      bool allowProdOfVars = true,
      bool allowExp = true,
      bool allowLog = true,
      bool allowInv = true,
      bool allowMultipleTerms = false
      ) {
      return new State(problemData, randSeed, maxVariables, scaleVariables, constOptIterations, lambda,
        collectParameterOptimalModels,
        lowerEstimationLimit, upperEstimationLimit,
        allowProdOfVars, allowExp, allowLog, allowInv, allowMultipleTerms);
    }

    // returns the quality of the evaluated solution
    public static double MakeStep(IState state) {
      var mctsState = state as State;
      if (mctsState == null) throw new ArgumentException("state");
      if (mctsState.Done) throw new NotSupportedException("The tree search has enumerated all possible solutions.");

      return TreeSearch(mctsState);
    }
    #endregion

    private static double TreeSearch(State mctsState) {
      var automaton = mctsState.automaton;
      var tree = mctsState.tree;
      var eval = mctsState.evalFun;
      var rand = mctsState.random;
      double q = 0;
      bool success = false;
      do {

        automaton.Reset();
        success = TryTreeSearchRec2(rand, tree, automaton, eval, mctsState, out q);
        mctsState.totalRollouts++;
      } while (!success && !tree.Done);
      if (success) {
        mctsState.effectiveRollouts++;

#if DEBUG
        Console.WriteLine(mctsState.ExprStr(automaton));
#endif

        return q;
      } else return 0.0;
    }

    // search forward
    private static bool TryTreeSearchRec2(IRandom rand, Tree tree, Automaton automaton,
      Func<byte[], int, double> eval,
      State state,
      out double q) {
      // ROLLOUT AND EXPANSION
      // We are navigating a graph (states might be reached via different paths) instead of a tree.
      // State equivalence is checked through ExprHash (based on the generated code through the path).

      // We switch between rollout-mode and expansion mode
      // Rollout-mode means we are navigating an existing path through the tree (using a rollout policy, e.g. UCB)
      // Expansion mode means we expand the graph, creating new nodes and edges (using an expansion policy, e.g. shortest route to a complete expression)
      // In expansion mode we might re-enter the graph and switch back to rollout-mode
      // We do this until we reach a complete expression (final state)

      // Loops in the graph are prevented by checking that the level of a child must be larger than the level of the parent
      // Sub-graphs which have been completely searched are marked as done.
      // Roll-out could lead to a state where all follow-states are done. In this case we call the rollout ineffective.

      while (!automaton.IsFinalState(automaton.CurrentState)) {
        // Console.WriteLine(automaton.stateNames[automaton.CurrentState]);
        if (state.children.ContainsKey(tree)) {
          if (state.children[tree].All(ch => ch.Done)) {
            tree.Done = true;
            break;
          }
          // ROLLOUT INSIDE TREE
          // UCT selection within tree
          int selectedIdx = 0;
          if (state.children[tree].Count > 1) {
            selectedIdx = SelectInternal(state.children[tree], rand);
          }

          tree = state.children[tree][selectedIdx];

          // move the automaton forward until reaching the state
          // all steps where no alternatives could be taken immediately (without expanding the tree)
          // TODO: simplification of the automaton 
          int[] possibleFollowStates = new int[1000];
          int nFs;
          automaton.FollowStates(automaton.CurrentState, ref possibleFollowStates, out nFs);
          Debug.Assert(possibleFollowStates.Contains(tree.state));
          automaton.Goto(tree.state);
        } else {
          // EXPAND 
          int[] possibleFollowStates = new int[1000];
          int nFs;
          string actionString = "";
          automaton.FollowStates(automaton.CurrentState, ref possibleFollowStates, out nFs);

          if (nFs == 0) {
            // stuck in a dead end (no final state and no allowed follow states)
            tree.Done = true;
            break;
          }
          var newChildren = new List<Tree>(nFs);
          state.children.Add(tree, newChildren);
          for (int i = 0; i < nFs; i++) {
            Tree child = null;
            // for selected states (EvalStates) we introduce state unification (detection of equivalent states)
            if (automaton.IsEvalState(possibleFollowStates[i])) {
              var hc = Hashcode(automaton);
              hc = ((hc << 5) + hc) ^ (ulong)tree.state; // TODO fix unit test for structure enumeration
              if (!state.nodes.TryGetValue(hc, out child)) {
                // Console.WriteLine("New expression (hash: {0}, state: {1})", Hashcode(automaton), automaton.stateNames[possibleFollowStates[i]]);
                child = new Tree() {
                  state = possibleFollowStates[i],
                  expr = actionString + automaton.GetActionString(automaton.CurrentState, possibleFollowStates[i]),
                  level = tree.level + 1
                };
                state.nodes.Add(hc, child);
              }
              // only allow forward edges (don't add the child if we would go back in the graph)
              else if (child.level > tree.level) {
                // Console.WriteLine("Existing expression (hash: {0}, state: {1})", Hashcode(automaton), automaton.stateNames[possibleFollowStates[i]]);
                // whenever we join paths we need to propagate back the statistics of the existing node through the newly created link
                // to all parents 
                BackpropagateStatistics(tree, state, child.visits);
              } else {
                // Console.WriteLine("Cycle (hash: {0}, state: {1})", Hashcode(automaton), automaton.stateNames[possibleFollowStates[i]]);
                // prevent cycles
                Debug.Assert(child.level <= tree.level);
                child = null;
              }
            } else {
              child = new Tree() {
                state = possibleFollowStates[i],
                expr = actionString + automaton.GetActionString(automaton.CurrentState, possibleFollowStates[i]),
                level = tree.level + 1
              };
            }
            if (child != null)
              newChildren.Add(child);
          }

          if (!newChildren.Any()) {
            // stuck in a dead end (no final state and no allowed follow states)
            tree.Done = true;
            break;
          }

          foreach (var ch in newChildren) {
            if (!state.parents.ContainsKey(ch)) {
              state.parents.Add(ch, new List<Tree>());
            }
            state.parents[ch].Add(tree);
          }


          // follow one of the children 
          tree = SelectStateLeadingToFinal(automaton, tree, rand, state);
          automaton.Goto(tree.state);
        }
      }

      bool success;

      // EVALUATE TREE
      if (!tree.Done && automaton.IsFinalState(automaton.CurrentState)) {
        tree.Done = true;
        tree.expr = state.ExprStr(automaton);
        byte[] code; int nParams;
        automaton.GetCode(out code, out nParams);
        q = eval(code, nParams);
        success = true;
        BackpropagateQuality(tree, q, state);
      } else {
        // we got stuck in roll-out (not evaluation necessary!)
        q = 0.0;
        success = false;
      }

      // RECURSIVELY BACKPROPAGATE RESULTS TO ALL PARENTS
      // Update statistics
      // Set branch to done if all children are done.
      BackpropagateDone(tree, state);
      BackpropagateDebugStats(tree, q, state);


      return success;
    }

    private static int SelectInternal(List<Tree> list, IRandom rand) {
      // choose a random node.
      Debug.Assert(list.Any(t => !t.Done));

      var idx = rand.Next(list.Count);
      while (list[idx].Done) { idx = rand.Next(list.Count); }
      return idx;
    }

    // backpropagate existing statistics to all parents
    private static void BackpropagateStatistics(Tree tree, State state, int numVisits) {
      tree.visits += numVisits;

      if (state.parents.ContainsKey(tree)) {
        foreach (var parent in state.parents[tree]) {
          BackpropagateStatistics(parent, state, numVisits);
        }
      }
    }

    private static ulong Hashcode(Automaton automaton) {
      byte[] code;
      int nParams;
      automaton.GetCode(out code, out nParams);
      return (ulong)ExprHashSymbolic.GetHash(code, nParams);
    }

    private static void BackpropagateQuality(Tree tree, double q, State state) {
      tree.visits++;
      // TODO: q is ignored for now

      if (state.parents.ContainsKey(tree)) {
        foreach (var parent in state.parents[tree]) {
          BackpropagateQuality(parent, q, state);
        }
      }
    }

    private static void BackpropagateDone(Tree tree, State state) {
      if (state.children.ContainsKey(tree) && state.children[tree].All(ch => ch.Done)) {
        tree.Done = true;
        // children[tree] = null; keep all nodes
      }

      if (state.parents.ContainsKey(tree)) {
        foreach (var parent in state.parents[tree]) {
          BackpropagateDone(parent, state);
        }
      }
    }

    private static void BackpropagateDebugStats(Tree tree, double q, State state) {
      if (state.parents.ContainsKey(tree)) {
        foreach (var parent in state.parents[tree]) {
          BackpropagateDebugStats(parent, q, state);
        }
      }

    }

    private static Tree SelectStateLeadingToFinal(Automaton automaton, Tree tree, IRandom rand, State state) {
      // find the child with the smallest state value (smaller values are closer to the final state)
      int selectedChildIdx = 0;
      var children = state.children[tree];
      Tree minChild = children.First();
      for (int i = 1; i < children.Count; i++) {
        if (children[i].state < minChild.state)
          selectedChildIdx = i;
      }
      return children[selectedChildIdx];
    }

    // scales data and extracts values from dataset into arrays
    private static void GenerateData(IRegressionProblemData problemData, bool scaleVariables, IEnumerable<int> rows,
      out double[][] xs, out double[] y, out double[] scalingFactor, out double[] scalingOffset) {
      xs = new double[problemData.AllowedInputVariables.Count()][];

      var i = 0;
      if (scaleVariables) {
        scalingFactor = new double[xs.Length + 1];
        scalingOffset = new double[xs.Length + 1];
      } else {
        scalingFactor = null;
        scalingOffset = null;
      }
      foreach (var var in problemData.AllowedInputVariables) {
        if (scaleVariables) {
          var minX = problemData.Dataset.GetDoubleValues(var, rows).Min();
          var maxX = problemData.Dataset.GetDoubleValues(var, rows).Max();
          var range = maxX - minX;

          // scaledX = (x - min) / range
          var sf = 1.0 / range;
          var offset = -minX / range;
          scalingFactor[i] = sf;
          scalingOffset[i] = offset;
          i++;
        }
      }

      if (scaleVariables) {
        // transform target variable to zero-mean
        scalingFactor[i] = 1.0;
        scalingOffset[i] = -problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows).Average();
      }

      GenerateData(problemData, rows, scalingFactor, scalingOffset, out xs, out y);
    }

    // extract values from dataset into arrays
    private static void GenerateData(IRegressionProblemData problemData, IEnumerable<int> rows, double[] scalingFactor, double[] scalingOffset,
     out double[][] xs, out double[] y) {
      xs = new double[problemData.AllowedInputVariables.Count()][];

      int i = 0;
      foreach (var var in problemData.AllowedInputVariables) {
        var sf = scalingFactor == null ? 1.0 : scalingFactor[i];
        var offset = scalingFactor == null ? 0.0 : scalingOffset[i];
        xs[i++] =
          problemData.Dataset.GetDoubleValues(var, rows).Select(xi => xi * sf + offset).ToArray();
      }

      {
        var sf = scalingFactor == null ? 1.0 : scalingFactor[i];
        var offset = scalingFactor == null ? 0.0 : scalingOffset[i];
        y = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows).Select(yi => yi * sf + offset).ToArray();
      }
    }

    // for debugging only


    private static string TraceTree(Tree tree, State state) {
      var sb = new StringBuilder();
      sb.Append(
@"digraph {
  ratio = fill;
  node [style=filled];
");
      int nodeId = 0;

      TraceTreeRec(tree, 0, sb, ref nodeId, state);
      sb.Append("}");
      return sb.ToString();
    }

    private static void TraceTreeRec(Tree tree, int parentId, StringBuilder sb, ref int nextId, State state) {
      var tries = tree.visits;

      sb.AppendFormat("{0} [label=\"{1}\"]; ", parentId, tries).AppendLine();

      var list = new List<Tuple<int, int, Tree>>();
      if (state.children.ContainsKey(tree)) {
        foreach (var ch in state.children[tree]) {
          nextId++;
          tries = ch.visits;
          sb.AppendFormat("{0} [label=\"{1}\"]; ", nextId, tries).AppendLine();
          sb.AppendFormat("{0} -> {1} [label=\"{2}\"]", parentId, nextId, ch.expr).AppendLine();
          list.Add(Tuple.Create(tries, nextId, ch));
        }

        foreach (var tup in list) {
          var ch = tup.Item3;
          var chId = tup.Item2;
          if (state.children.ContainsKey(ch) && state.children[ch].Count == 1) {
            var chch = state.children[ch].First();
            nextId++;
            tries = chch.visits;
            sb.AppendFormat("{0} [label=\"{1}\"]; ", nextId, tries).AppendLine();
            sb.AppendFormat("{0} -> {1} [label=\"{2}\"]", chId, nextId, chch.expr).AppendLine();
          }
        }

        foreach (var tup in list.OrderByDescending(t => t.Item1).Take(1)) {
          TraceTreeRec(tup.Item3, tup.Item2, sb, ref nextId, state);
        }
      }
    }

    private static string WriteTree(Tree tree, State state) {
      var sb = new System.IO.StringWriter(System.Globalization.CultureInfo.InvariantCulture);
      var nodeIds = new Dictionary<Tree, int>();
      sb.Write(
@"digraph {
  ratio = fill;
  node [style=filled];
");
      int threshold = /* state.nodes.Count > 500 ? 10 : */ 0;
      foreach (var kvp in state.children) {
        var parent = kvp.Key;
        int parentId;
        if (!nodeIds.TryGetValue(parent, out parentId)) {
          parentId = nodeIds.Count + 1;
          var tries = parent.visits;
          if (tries > threshold)
            sb.Write("{0} [label=\"{1}\"]; ", parentId, tries);
          nodeIds.Add(parent, parentId);
        }
        foreach (var child in kvp.Value) {
          int childId;
          if (!nodeIds.TryGetValue(child, out childId)) {
            childId = nodeIds.Count + 1;
            nodeIds.Add(child, childId);
          }
          var tries = child.visits;
          if (tries < 1) continue;
          if (tries > threshold) {
            sb.Write("{0} [label=\"{1}\"]; ", childId, tries);
            var edgeLabel = child.expr;
            // if (parent.expr.Length > 0) edgeLabel = edgeLabel.Replace(parent.expr, "");
            sb.Write("{0} -> {1} [label=\"{2}\"]", parentId, childId, edgeLabel);
          }
        }
      }

      sb.Write("}");
      return sb.ToString();
    }
  }
}