ferrall / niqlow Goto Github PK

If I set detergent.ox the maxiter property to

mle.maxiter = 15;

in the detergent.ox file of DCDPdetergent the file does not even complete one iteration. See the newest log file (bfgs_maxiter.log) in the gist

I created a new state variable class for my coupon variables and I now have Stage 1 of my estimator working.

Stage 2 if I set EMax.DoNotIterate = TRUE; then I get this log file

The file shows that cur->Outcome::FullLikelihood() = .NaN; which doesn't seem right.

If I set EMax.DoNotIterate = FALSE; I get the stack trace below

error while checking convergence
Runtime error occurred in Update (212), call trace:
 (212): Update
 (549): loop
 (625): Traverse
 (123): Gsolve
 (85): Run
 (229): loop
 (66): Run
 (166): Solve
 (503): LogLikelihood
 (531): LogLikelihood
 (836): EconometricObjective

which is also due to a .NaN value.

Any debugging tips?

Current issues:

1. Solver not converging when I put the consumption FixedState in Endogenous(...)

   Amoeba method of NelderMead gets stuck at a plex size of 0.2 at a variety of starting values. Also tried BFGS and got

   Runtime error: '[-1] in array[4]' index out of range
   Runtime error occurred in Utility (103), call trace:
   detergent.ox (103): Utility
   

   is there anything wrong with my Utility method?

2. Putting FixedEffect variable in GroupVariables(...) does not seem to work

   Using NelderMead I get

   Plex size tolerance 1e-05
   Amoeba Start
   now: 0
   viinds:
   [0] =
        2.0000
   [1] =
        3.0000
   
   vilikes: 0
   Runtime error: 'matrix[0][0]' cannot cast to double
   Runtime error occurred in PathLike (404), call trace:
   (404): PathLike
   (447): Likelihood
   (515): LogLikelihood
   (532): LogLikelihood
   (862): EconometricObjective
   (550): vfunc
   (345): vobj
   (286): funclist
   (333): Amoeba
   

   using a modified PathLike() method in Outcomes.ox

   /** Integrate over the path.
   
   **/
   Path::PathLike() {
     decl cur, glk;
     now = TRUE;
     print("now: ", !now, "\n");
     print("viinds: ", viinds, "\n");
     print("vilikes: ", vilikes, "\n");
     viinds[!now] = <>;
     vilikes[!now] = <>; // needs to be array - isn't
     print("complete");
     cur = last;
     do {
         tom = !now;
         cur->Outcome::Likelihood();
         now = !now;
         } while(isclass(cur = cur.prev));
     L = double(sumr(vilikes[!now]));
     }
   

   vilikes is an integer and it needs to be a matrix

3. How long should I expect estimating the model to take if the state space is around 10000 and there are about 1000 households?

4. Do you know of a way to track the progress of the model on sharcnet (using logging)? When I submit a job on sharcnet I don't get any results until the model times out, completes or crashes because it doesn't write the output to a file until the very end.

I am still having a problem with histograms in the PanelPrediction output not adding to 1 (though the new average print-out helps). I have replaced the old output file in our branch. Thanks.

I tried estimating Stage 2 (only) of my model and get the following error

Runtime error: 'pandv' member function call requires object
Runtime error occurred in GetPstar (54), call trace: 
 (54): GetPstar
 (376): FullLikelihood
 (447): FullLikelihood
 (508): LogLikelihood
 (525): LogLikelihood
 (859): EconometricObjective
 (555): vfunc
 (350): vobj
 (291): funclist
 (404): Jacobian
 (376): Gradient
remaining calls not listed

The full log file is here. Any idea what's wrong? My model code is on DCDPdetergent.git in the peanut_butter branch.

The dataset is in my sharcnet directory /home/cpritcha/paper/src/data/pbutter.dta

I wrote a simplified version of the model to find a few mistakes (I've shared it with you in the github sept directory).

If I set the RandomUpDown to just simply stay constant "(010)", then the predicted histograms add up to 1 until I reach the stationary last period (T-2) at which point the choice probabilities and distributions do not make sense. As soon as I change the RandomUpDown variable so that it increases or decreases with some probability, I begin to "lose" people in predicted distributions in the "age phase" of the model.

Do I need to set up the RandomUpDown probabilities so it checks the current state to see if it is at the max or min value and subsequently adjusts the probabilities to ensure the state variable does not go above the set max value?

Also: I'm not sure how to interpret the transitions when I set the volume to noisy.