Support for time lags in constant optimizer

[gkronber]

14349: added support for lagged variables to constants optimizer

[gkronber]

r14358: in the constants optimizer the variable weight must be multiplied to the variable value even when variable weights are not updated by the constants optimizer

[mkommend]

TestScript for constant optimization

[mkommend]

ConstantOptimization of lagged variables does not work! The issue is that the symbol LaggedVariable is derived from Variable and as a result of using the is operator (check if the type is in the inheritance hierarchy) all lags are always 0. The if block in line 311 is never executed.

Any easy solution would be to switch the if blocks for variable and lagged variable. I haven't performed this change due to refactoring of the CoOp / AutoDiff transformation.

The attached script CoOp Test also illustrates this bug. In the current implementation a R² value of 1 could not be reached and the constants are tuned incorrectly. If the update is performed 0.999999999... is reached.

[mkommend]

Reviewed r14349 and r14358.

Thanks for r14358. This was a bug in the CoOp.

[gkronber]

Replying to mkommend:

ConstantOptimization of lagged variables does not work! The issue is that the symbol LaggedVariable is derived from Variable and as a result of using the is operator (check if the type is in the inheritance hierarchy) all lags are always 0. The if block in line 311 is never executed.

This should be fixed once the factors-branch (#2650) is integrated. After integration of the factors branch this problem should be tested again.

[gkronber]

Replying to gkronber:

Replying to mkommend:

ConstantOptimization of lagged variables does not work! The issue is that the symbol LaggedVariable is derived from Variable and as a result of using the is operator (check if the type is in the inheritance hierarchy) all lags are always 0. The if block in line 311 is never executed.

This should be fixed once the factors-branch (#2650) is integrated. After integration of the factors branch this problem should be tested again.

I just checked this again and it works now.

[gkronber]

Please set ticket to ready for release if there are no further issues.

[mkommend]

Replying to gkronber:

Replying to gkronber:

Replying to mkommend:

ConstantOptimization of lagged variables does not work! The issue is that the symbol LaggedVariable is derived from Variable and as a result of using the is operator (check if the type is in the inheritance hierarchy) all lags are always 0. The if block in line 311 is never executed.

This should be fixed once the factors-branch (#2650) is integrated. After integration of the factors branch this problem should be tested again.

I just checked this again and it works now.

This is not true. The script still does not give an R² value of ~ 1 when variable weights are tuned.

[mkommend]

r14946: Fixed updating of constants / variable weights if lagged variables are used within constants optimization.

[gkronber]

Thanks for re-checking & fixing this.

[mkommend]

r14951: Code clean up in ConstantOptimization.

[gkronber]

r14962 merged r14349 and r14358 from trunk to stable (NOTE: note completely merged to stable yet)

Remaining changes should be merged after r14826 in #2650.

[gkronber]

r15136: merged r14946 and r14951 from trunk to stable (all changesets merged)