Pruning of introns/negative impact nodes in symbolic data analysis expressions

[bburlacu]

An operator/analyzer should be implemented that calculates the impacts of individual nodes in the solution individuals and removes the ones with an impact value below a certain user-defined threshold. The user should also be able to choose between only removing introns (nodes with zero impact) or nodes with a negative impact as well.

[bburlacu]

r10368: Implemented symbolic data analysis pruning operator and analyzers.

[bburlacu]

r10375: Fixed incorrect accessibility levels of constructors for the pruning analyzers.

[bburlacu]

r10378: Added storable constructors.

[bburlacu]

r10414: Modified the pruning operator and analyzer to use the FitnessCalculationPartition for impact and replacement values calculation, instead of the whole training data partition.

[bburlacu]

r10417: SymbolicDataAnalysisExpressionPruningOperator: changed cloning constructor accessibility to protected. Added storable constructor.

[bburlacu]

r10418: Added license header to SymbolicClassificationPruningAnalyzer.cs

[bburlacu]

r10428: Reset the pruned subtrees/trees counters to zero on reinitialization.

[mkommend]

Reviewing comments:

• SymbolicDataAnalysisSingleObjectivePruningAnalyzer
  • missing license headers
  • Don't abuse the StatefulItem methods for counters, use global scope variables instead
  • Use properties instead of variables for storable members (e.g., impact values caluclator)
  • Program as defensive as possible - use readonly attribute
  • Every value parameter must have a non-empty description
  • Use only one operator instance instead of creating multiple ones and subscopes processor
  • Use a DataTableValuesCollector for creating the DataTables
  • Use the mechansim of the MultiAnalyzer for checking the update interval
  • Returning a correctly configured operation collection will get rid of reentering the analyzer

• SymbolicDataAnalysisExpressionPruningOperator
  • Use Lookup- and ValueParameters
  • Evaluation happens multiple times and should be unified, consider changing the impact value calculators
  • Do not use locks
  • Random property is not used at all
  • Move the model creation in the pruning operator

[bburlacu]

r10469: Refactored pruning analyzer and operators as per review.

[bburlacu]

r10470: Fixed deep cloning error and assembly plugin reference.

[bburlacu]

r10955: Added extra check to ensure the PopulationSlice parameter is in the right range.

[bburlacu]

r11013: Added missing check for population slice bounds, added PopulationSize parameter to get the number of individuals (instead of counting the subscopes of the current execution context).

[mkommend]

Common things

• missing license header
• no properties for lookup parameters
• parameters for fixed value parameters should provide getter and setter for the standard c# data types (e.g IFixedValueParameter<BoolValue> xParam => bool x {get{return xParam.Value}set {xParam.Value = value}})
• empty lines between methods
• avoid ToList in code for fitness calculation
• unused properties and fields

SymbolicDataAnalysisExpressionPruningOperator

• unify parameter descriptions
• abstract methods should be below the ctors to make them more prominent

• Quality of trees is not updated
• Fitness calculation partition is used for impact calculation but the whole training partition is used for model evaluation !!!
• Try to use ToList as rarely as possible, e.g. it is not necessary for the rows
• Do not create a new Constant symbol, get the constant symbol from the grammar so that it has the correct local parameters

Test if any nodes are pruned if PruneOnlyZeroImpactNodes is active

SymbolicRegressionPruningOperator

• avoid ToList
• use the parameter property instead to access lookup parameter values
• Evaluator parameter is not used => remove the parameter

Test if existing experiments can still be opened

SymbolicClassificationPruningOperator

• ApplyLinearScalingParameter is never used => remove the parameter + property
• Remove property for the model creator
• CreateModel should use the whole training data
• Evaluate should use only the fitness calculation partition
• Avoid ToList
• the classification is quality is given in terms of the accuracy and not the R²
• License header is missing

SymbolicDataAnalysisSingleObjectivePruningAnalyzer

• move initialize operators to the end of the file
• UpdateCounter, UpdateInterval, PruningProbability, PopulationSliceParameter should be FixedValueParameters
• empty lines between regions
• rework parameter descriptions
• rename oc to operations or something similar and more descriptive
• initialize the necessary operator in the ctor
• remove property for the emtpy operator
• do not create the operations for data collection by yourself => chain the successor parameters of the analyzer and the operators and let the engine do the work => base apply call must be inserted and the cloning won't be necessary

[bburlacu]

r11025: Attempted to fix the problems described above.

[bburlacu]

r11026:

• Removed ToArray() calls from SymbolicClassificationPruningOperator.cs and SymbolicRegressionPruningOperator.cs.
• Changed UpdateCounter, UpdateInterval, PruningProbability, PopulationSliceParameter to FixedValueParameters in SymbolicDataAnalysisSingleObjectivePruningAnalyzer.cs.
• Removed empty operator data member
• Added parameter descriptions
• Improved source organization

[bburlacu]

r11027: SymbolicDataAnalysisSingleObjectivePruningAnalyzer.cs: Added some code in the AfterDeserialization hook to replace ValueParameters with FixedValueParameters in order to avoid an InvalidCastException when trying to run older saved algorithms.

[mkommend]

Reviewed r11025, r11026 & r11027.

[mkommend]

r11145: Merged r10368, r10375, r10378, r10414, r10417, r10418, r10428, r10469, r10470, r11013, r11025, r11026, r11027 into stable.