source: branches/3022-FastFunctionExtraction/FFX/FFXModel.cs @ 17779

using HeuristicLab.Algorithms.DataAnalysis.Glmnet;
using HeuristicLab.Problems.DataAnalysis;
using HeuristicLab.Problems.DataAnalysis.Symbolic;
using HeuristicLab.Problems.DataAnalysis.Symbolic.Regression;
using System;
using System.Collections.Generic;
using System.Globalization;
using System.Linq;

namespace HeuristicLab.Algorithms.DataAnalysis.FastFunctionExtraction {
    internal class FFXModel {
        public FFXModel(double intercept, IEnumerable<(double coeff, IBasisFunction function)> basisFunctions) {
            Intercept = intercept;
            BasisFunctions = basisFunctions ?? throw new ArgumentNullException(nameof(basisFunctions));
        }

        public double Intercept { get; set; }
        public IEnumerable<(double coeff, IBasisFunction function)> BasisFunctions { get; set; }
        public IEnumerable<(double coeff, IBasisFunction function)> NominatorFunctions =>
            BasisFunctions.Where(bf => bf.function.IsDenominator);
        public IEnumerable<(double coeff, IBasisFunction function)> DenominatorFunctions =>
            BasisFunctions.Where(bf => !bf.function.IsDenominator);

        public int NumNumeratorFunctions => NominatorFunctions != null ? NominatorFunctions.Count() : 0;
        public int NumDenominatorFunctions => DenominatorFunctions != null ? DenominatorFunctions.Count() : 0;
        public int NumBases => NumNumeratorFunctions + NumDenominatorFunctions;
        public IEnumerable<IBasisFunction> Bases =>
            NominatorFunctions.Select(tuple => tuple.function)
            .Concat(DenominatorFunctions.Select(tuple => tuple.function));
        public int Complexity {
            get {
                // We have a leading constant, then for each base we have a coefficient, 
                // a multiply, and a plus, plus the complexity of the base itself.
                int numerator_complexity = 1 + NominatorFunctions.Select(bf => 3 + bf.function.Complexity).Sum();
                if (NumDenominatorFunctions == 0) {
                    return numerator_complexity;
                } else {
                    int denominator_complexity = 1 + DenominatorFunctions.Select(bf => 3 + bf.function.Complexity).Sum();
                    return 1 + numerator_complexity + denominator_complexity;
                }
            }
        }

        public override string ToString() {
            var culture = new CultureInfo("en-US");
            var result = Intercept.ToString(culture);
            foreach (var (coeff, function) in NominatorFunctions) {
                var str = $" + ({Math.Round(coeff, 4).ToString(culture)}) * {function}";
                result += str;
            }
            if (NumDenominatorFunctions == 0) return result;

            result = "(" + result + ") / (1.0";
            foreach (var (coeff, function) in DenominatorFunctions) {
                var str = $" + ({coeff.ToString(culture)}) * {function}";
                result += str;
            }
            result += ")";
            return result;
        }

        public double[] Simulate(IRegressionProblemData data, IEnumerable<int> rows) {
            var symbolicRegressionModel = ToSymbolicRegressionModel(data.TargetVariable);
            return symbolicRegressionModel.GetEstimatedValues(data.Dataset, rows).ToArray();
        }

        public ISymbolicRegressionModel ToSymbolicRegressionModel(string targetVariable) {
            var str = this.ToString();
            var tree = new InfixExpressionParser().Parse(str);
            return new SymbolicRegressionModel(targetVariable, tree, new SymbolicDataAnalysisExpressionTreeInterpreter());
        }

        public void OptimizeCoefficients(IRegressionProblemData data) {
            var elnetData = BFUtils.PrepareData(data, this.BasisFunctions.Select(tuple => tuple.function));
            var oldCoeff = this.BasisFunctions.Select(tuple => tuple.coeff).ToArray();
            var oldIntercept = this.Intercept;
            var newCoeff = ElasticNetLinearRegression.CalculateModelCoefficients(elnetData, 0, 0, out var _, out var _);
            var newIntercept = newCoeff.Last();
            newCoeff = newCoeff.Take(newCoeff.Length - 1).ToArray();

            var bfArray = this.BasisFunctions.ToArray();
            for (int i = 0; i < bfArray.Length; i++) {
                bfArray[i] = (newCoeff[i], bfArray[i].function);
            }

            this.BasisFunctions = bfArray;
            this.Intercept = newIntercept;
        }
    }
}