Implement Intervals and according calculation rules

[mkommend]

This ticket collects the changes for the implementation of intervals and calculation rules and acts as base for a-priori knowledge integration and further model analysis. The following class will be added in this ticket:

• Interval
• Interpreter for symbolic models to calculate intervals of a model
• Unit test for the calculation with intervals

[chaider]

r16320: Created Branch with problems data analysis and problems data analysis symbolic.

[chaider]

r16322: Added references to projects and added build events.

[chaider]

r16323: Added Interval class and Interpreter class.

[chaider]

r16324: Added test class to solution.

[mkommend]

r16326: Corrected namespace of intervals.

[mkommend]

r16327: Changed bounds to auto property in Interval and corrected typo in comment.

[mkommend]

r16328: Reordered methods in interval interpreter, added IStatefulItem interface, sealed class.

[mkommend]

Review comments

Interval

• Everything looks file

IntervalInterpreter

• Source file is missing an I
• Item Description must be improved. It reads currently "Interpreter for interval arithmetic within symbolic regression.", which is only partially true, because this is an interpreter for calculation of intervals of arbitrary symbolic models (e.g., it could be also used for linear regression).
• Line 46: What is the purpose of GetSymbolicExpressionTreeIntervals(tree, rows, out intervals)? How can this method calculate anything if no ranges for inputs or a dataset is provided?
• Why does each method for interval calculation has a row parameter? What is the purpose of the parameter? Is it really beneficial to calculate an interval of a dataset only for parts of it?
• Why is there no method GetSymbolicExpressionTreeInterval without an out parameter?
• What is the purpose of the syncRoot if it is not used at all?
• Line 77: move declaration of unmodified varialbe necessaryArgStackSize closer to its usage in Line 112. Perhaps define it as a const.
• Is it not possible to get rid of the interpreter state and just use an array of instructions?
• Refactor assignment conditionals for interval calculation
  • if(customIntervals[varName] != null ...
• Implement null check for dataset and customIntervals as preconditions at the start of an method
• Check for variable occurrences once to avoid the containskey statements every time
• Evaluate should have the intervals parameter as optional parameter and only if it is assigned the intermediate intervals should be calculated

Feel free to ask questions if anything is unclear.

[chaider]

r16330: Implemented review comments and added IsCompatible method.

[chaider]

r16331: Changed Evaluate method (changed intervals parameter to optional)

[mkommend]

Review Comments

IntervalInterpreter

• Line 142 throws an Exception instead of an ArgumentException
• Line 144 fails if the Dataset is null, which is not checked
• Line 152 checks for custom intervals however no error handling is implemented if the specific custom interval is not present (line 153).
• Extraction of intervals for variables of the dataset should be done only once; in contrast to each occurrence of the variable.

[chaider]

r16363 Added GetVariableBoundaries method in DatasetUtil

[chaider]

r16364

• Changed signature of GetSymbolicExressionTreeIntervals methods
• Changed PrepareInterpreterState (removed optinal parameters, takes Dictionary<string, Interval> as input and no dataset anymore)
• Added optional parameter (rows) to GetVariableBoundaries method (allows to use just parts from dataset as input e.g. training/test indices)

[chaider]

r16367 Added EvaluatedSolutions Parameter

[chaider]

r16374 renamed variable from intermediate to result

[chaider]

r16376 Removed member variable InstructionCount and added ref parameter to Evaluate method instead

[chaider]

r16377 Changed namespace to HeuristicLab.Problems.DataAnalysis.Symbolic

[chaider]

r16380 added variable weight multiplication to the evaluate method

[chaider]

r16383

• Some renaming and reordering
• Changed GetVariableRanges method in DatasetUtil
• Added cases for substract and divide with arity 1

[chaider]

r16393 Checked null references

[chaider]

r16403 Changed variable names

[chaider]

r16404

• Variable name changes
• Changed GetVaribleRanges in DatasetUtil
• Added method to get interval ranges from a double enumeration in Interval class
• Added some comments

[chaider]

r16406 Changed isEqual method in Interval

[chaider]

r16407: Merged branch changes into trunk.

[chaider]

r16409: deleted branch

[mkommend]

r16436: Merged r16407 into stable.

[gkronber]

The interpreter might produce invalid intervals where the lower bound is larger than upper bound.

[gkronber]

r16629: added extra checks to make sure that IntervalInterpreter always returns a valid interval

[gkronber]

r16631: fixed calculation for intervals of Sqr, Sqrt, Cube, CubeRoot

[gkronber]

r16646: fixed bug in Interval.Multiply

[gkronber]

Interval interpreter does not support factor variables. Test with Miba problem instances.

[gkronber]

There are several changes made in the context of this ticket which have not been merged to stable after r16436.

[gkronber]

r16740: fixed method names

[gkronber]

r16743: also fixed method names in unit tests

[gkronber]

The interval evaluator does not yet support tanh (see #2866)

[gkronber]

Since interval arithmetic automatically evaluates all solutions it must support all symbols, otherwise algorithms cannot be executed (are stopped because of exceptions).

[gkronber]

r16757: fixed bug in interval calculation for cos(x) and added/fixed unit tests for sin and cos

[gkronber]

r16758: added support for tanh to interval interpreter and added unit tests for tan and tanh

[gkronber]

depends on #2866

[gkronber]

r16769: added error-thresholds for comparisons in unit tests (hopefully fixing tests)

[gkronber]

r16822: changed method signatures in IntervalInterpreter to use interface IDictionary instead of Dictionary

[mkommend]

r17100: Merged 16629, 16631, 16646, 16740, 16743, 16757, 16758, 16769, 16822 to stable.

[gkronber]

r17167: added info that r16407 has been merged to stable (see r16436)