Opened 5 months ago

Last modified 5 weeks ago

#2974 new enhancement

Refactor Constants Optimization

Reported by: mkommend Owned by: mkommend
Priority: medium Milestone: HeuristicLab 3.3.17
Component: Problems.DataAnalysis.Symbolic.Regression Version: branch
Keywords: Cc:

Description

The code for constants optimization has grown over the years and all new functionality has been directly added in the evaluator. This should be refactored into separate classes and improvements to its functionality should be implemented. In detail the following changes should be realized:

  • Optimizing constants should be separated from evaluation
  • Constant optimization should use the MSE instead of the R²
  • Add interfaces and static utility methods
  • ...

Attachments (1)

benchmark_16691.png (27.8 KB) - added by gkronber 2 months ago.

Download all attachments as: .zip

Change History (24)

comment:1 Changed 5 months ago by mkommend

r16455: Created branch and added adapted project HL.Problems.DataAnalysis.Symbolic.Regression.

comment:2 Changed 5 months ago by mkommend

r16456: Added HL.Problems.DataAnalysis.Symbolic to the branch.

comment:3 Changed 5 months ago by mkommend

r16457: Extracted linear scaling terms in auto diff converter.
r16458: Readded accidently remove readonly modifier in autodiff converter.

comment:4 Changed 5 months ago by mkommend

r16459: Added utility methods and copy of constant optimization evluator for refactoring.
r16460: Adapted project files.

Last edited 5 months ago by mkommend (previous) (diff)

comment:5 Changed 5 months ago by mkommend

r16461: Added unit tests and refactoring.

comment:6 Changed 5 months ago by mkommend

r16463: Adapted tree to autodiff converter.
r16464: Adapted new constants optimizer.

Last edited 5 months ago by mkommend (previous) (diff)

comment:7 Changed 5 months ago by mkommend

r16500: Added intermediate version of new constants optimization for profiling.

comment:8 Changed 4 months ago by mkommend

r16507: First stable version of new CoOp.
r16508: Reverted changes in TreeToAutoDiffTermConverter.

comment:9 Changed 4 months ago by mkommend

r16509: Added new evaluator with constants optimization.

comment:10 Changed 4 months ago by mkommend

r16514: Updated CoOp to handle negative R² values approprietly.

comment:11 Changed 4 months ago by mkommend

r16522: Fixed bug in constants extraction when adding linear scaling coefficients.

comment:12 Changed 4 months ago by bburlacu

r16525: Add benchmarks solution for testing constant optimization performance.

comment:13 Changed 2 months ago by gkronber

r16676: merged r16478:16672 from trunk:HeuristicLab.Problems.DataAnalysis.Symbolic to branch:HeuristicLab.Problems.DataAnalysis.Symbolic

comment:14 Changed 2 months ago by gkronber

r16677: merged r16499:16658 from trunk:HeuristicLab.Problems.DataAnalysis.Symbolic.Regression to branch:HeuristicLab.Problems.DataAnalysis.Symbolic.Regression

comment:15 Changed 2 months ago by gkronber

r16678: removed project references again after merging changes from trunk

comment:16 Changed 2 months ago by gkronber

r16679: changed .NET framework version of UnitTests project for compatibility with HL trunk (4.6.1)

comment:17 Changed 2 months ago by gkronber

r16680: necessary changes for compatibility with new persistence

comment:18 Changed 2 months ago by gkronber

r16689: changed LMConstantsOptimizer to use minlm interface of alglib instead of lsfit interface.

comment:19 Changed 2 months ago by gkronber

r16690: added code in the unit tests to make behaviour of new and old ConstOpt consistent to allow direct comparison of results (+fixed a small bug in the unit tests)

comment:20 Changed 2 months ago by gkronber

Currently, one unit test fails because of code in OnlinePearsonRCalculator.

It would work if we would replace it with

    public double R {
      get {
        double xVar = sxCalculator.PopulationVariance;
        double yVar = syCalculator.PopulationVariance;
        var r = covCalculator.Covariance / (Math.Sqrt(xVar) * Math.Sqrt(yVar));
        if (double.IsNaN(r)) return 0.0;
        if (r < -1.0) r = -1.0;
        else if (r > 1.0) r = 1.0;
        return r;
      }
    }

comment:21 Changed 2 months ago by gkronber

r16691: improved code for benchmarking (previous code failed to account for update of trees after call of Optimize as well as early stopping in Optimize for trees with invalid outputs)

comment:22 Changed 2 months ago by gkronber

Benchmark results on my notebook (10 trees, maxSize: 50, Tower)

Last edited 2 months ago by gkronber (previous) (diff)

Changed 2 months ago by gkronber

comment:23 Changed 5 weeks ago by gkronber

  • Milestone changed from HeuristicLab 3.3.16 to HeuristicLab 3.3.17
Note: See TracTickets for help on using tickets.