AR1 is not evaluated correctly on the test partition

[gkronber]

The AR1 model should regress on the previous model outputs in the test set (therefore the name auto-regressive).

Currently the AR1 model uses the actual value of the target variable at t-1 even on the test partition.

This is also true for the symbol "auto-regressive variable" for symbolic time-series modeling.

[gkronber]

Problematic code in AutoregressiveModeling.cs:

      for (int i = 0; i < timeOffset; i++) {
        LaggedVariableTreeNode node = (LaggedVariableTreeNode)new LaggedVariable().CreateTreeNode();
        node.VariableName = targetVariable;
        node.Weight = coefficients[i];
        node.Lag = (i + 1) * -1;
        addition.AddSubtree(node);
      }

The symbol AutoregressiveTargetVariable is treated in the same way as any other lagged variable in the interpreter

        case OpCodes.LagVariable: {
            var laggedVariableTreeNode = (LaggedVariableTreeNode)currentInstr.dynamicNode;
            int actualRow = row + laggedVariableTreeNode.Lag;
            if (actualRow < 0 || actualRow >= dataset.Rows) { return double.NaN; }
            return ((IList<double>)currentInstr.data)[actualRow] * laggedVariableTreeNode.Weight;
          }

[gkronber]

Attached file shows unwanted behavior.

[gkronber]

The caching of target variable values in the interpreter for time series models has a bug (the values are not initialized correctly e.g. when the problem changes).

A specific version of the line chart needs to be implemented which uses the complete test partition as prognosis horizon.

[gkronber]

r14422: added a way to calculate prognosed values for the whole test partition and added a specific implementation of the line chart for time series models

[gkronber]

r14423: removed caching of traget variable values because it didn't work correctly.

[gkronber]

TODO: check if sample still produces the same values.

[gkronber]

Replying to gkronber:

TODO: check if sample still produces the same values.

Nope...

Stacktraces: (new) MSTest: HeuristicLab.Tests.GPTimeSeriesSampleTest.RunGpTimeSeriesSampleTest Assert.AreEqual failed. Expected a difference no greater than <1E-08> between expected value <0,0209527534151996> and actual value <0,0490851910414951>.

[gkronber]

Merge with #2718

[gkronber]

r14529: updated target result values in unit test (previously the interpreter was buggy)

[mkommend]

OK. I still find it slightly misleading that the training and test range behave differently (estimation vs. prognosis).

[gkronber]

r14939: merged r14422 and r14423 from trunk to stable

r14941: merged r14529 from trunk to stable