Calculate the complexity of symbolic expression trees

[mkommend]

Currently the only measure that gives an indication of the complexity of an symbolic data analysis model is the length of the expression tree. Better estimators based on the the tree length or the functional complexity are needed to perform multi-objective optimization.

[mkommend]

Branch creation

r10748: Added branch folder for complexity analysis.
r10749: Branched DataAnalysis.Symbolic with local changes.
r10750: Branched DataAnalysis.Symbolic.Regression with local changes.
r10751: Added solution for complexity analysis.

[mkommend]

r11310: Merged trunk changes into complexity branch.

[mkommend]

r11407: Added nested tree size evaluator and simplification operator.

[mkommend]

r11861: Merged trunk changes.

[mkommend]

r11862: Commited additional files from trunk merge.

[mkommend]

r11863: Corrected accidentally commited type in tree simplifier.

[mkommend]

r11883: Additional calculation of results for EuroCast 2015.

[mkommend]

r12130: Updated complexity branch with trunk changes.

[mkommend]

r12147: Added rounding of quality values to multi-objective sym reg evaluators.

[mkommend]

r12187: Corrected error in complexity calculation for sine.

[mkommend]

r12214: Merged trunk changes into complexity branch.

[mkommend]

r12233: Merged trunk changes.

[mkommend]

r12848: Merged trunk changes and extracted parameters of evaluators to their base class.

[mkommend]

r12849: Updated license year to 2015.

[mkommend]

r12850: Forgot to commit merge info and some files in r12848.

[mkommend]

r13211: Added parameter for constant optimization iterations.

[mkommend]

r13213: Merged trunk changes into branch for complexity analysis.

[mkommend]

r13220: Improve source code and usability of complexity calculator.

[mkommend]

r13221: Removed unused features from complexity analysis.

[mkommend]

I recommend to perform the trunk integration locally and just review the changes in the working copy. Otherwise it could be quite cumbersome to identify all changes.

[gkronber]

r13241: merged all changes from branch to trunk (preserving the original .csproj-files so that I don't have to change reference and output paths)

[gkronber]

Review comments:

• Why is the tree simplified twice in the SymbolicDataAnalysisExpressionTreeSimplificationOperator?
• I think it is more readable if cases that are treated equally in the ComplexityCalculater have 'fall-through' cases in the switch statement.
• The calculation of the complexity for the power and root symbol should be explained (in a comment?)
• Why have properties TrainingBestSolutions and TrainingBestsolutionQualities in ...MultiObjectiveTrainingBestSolutionAnalyzer been changed from public to private? The values are accessible anyway through the parameter properties.

r13300: made some changes while reviewing the code r13307: fixed a typo while reviewing

[gkronber]

r13308: terminated feature branch (changesets have been merged to trunk)

[gkronber]

Release following changesets to stable before closing this ticket as done:

• r13241, r13300, r13307

[mkommend]

r13310: Merged r13241, r13300, r13307, r13308 into stable.

[mkommend]

Replying to gkronber:

Review comments:

• Why is the tree simplified twice in the SymbolicDataAnalysisExpressionTreeSimplificationOperator?

Because the simplifier does not perform all simplification in one pass. Hence, the simplifier is called twice.

• I think it is more readable if cases that are treated equally in the ComplexityCalculater have 'fall-through' cases in the switch statement.

I totally agree.

• The calculation of the complexity for the power and root symbol should be explained (in a comment?)

The calculation of the complexity of power and root have not been extensively tested, because I use mostly use square and square root while modeling instead of power and root.

• Why have properties TrainingBestSolutions and TrainingBestsolutionQualities in ...MultiObjectiveTrainingBestSolutionAnalyzer been changed from public to private? The values are accessible anyway through the parameter properties.

Because these access the actual value of the lookup parameters and are therefore only accessible when an execution context is set for the operator (exception otherwise). When I access those values through the ILookupParameter I am aware of this issue.

r13300: made some changes while reviewing the code r13307: fixed a typo while reviewing

Thanks