MeganRSullivan
commented
1 month ago When I tried skipping ResetPar entirely, the model crashed in the first iteration (after the initial iteration). The oxygen model never converged, and then the model crashed in neglogpost because: Output argument Ox (and maybe others) not assigned during call to eqOcycle_v2.
See log file below:
...
...
...
current objective function value is 1.947e+00
saving data to ../output/optPCO_constC2P_v2_CTL_He_PCO_DOC1_DOP0.mat ...
saving xhat to ../output/optPCO_constC2P_v2_CTL_He_PCO_DOC1_DOP0_xhat.mat ...
Norm of First-order
Iteration f(x) step optimality CG-iterations
0 1.94732 406
Tunable parameters
-----------------------------------------------
current sigP is 2.97e-03
current Q10P is 1.65e+00
current kdP is 2.40e-08
current bP_T is -4.47e+00
current bP is 5.47e+02
current alpha is 7.79e-08
current beta is 4.36e-01
current sigC is 3.59e-01
current kru is 3.64e-12
current krd is 4.90e-13
current etau is 1.03e+00
current bC_T is 1.93e+00
current bC is 1.74e+00
current d is 4.11e+03
current Q10C is 4.63e-01
current kdC is 1.18e-08
current R_Si is -7.34e-01
current rR is 1.65e-02
current dd is 1.38e-02
current rO2C is 1.24e+00
saving xhat & x0 to: ../output/optPCO_constC2P_v2_CTL_He_PCO_DOC1_DOP0_xhat.mat ...
current iteration is 1
Solving P model ...
factoring Jacobian matrix ...
Elapsed time is 340.027612 seconds.
solve for P-cycle model state ...
Elapsed time is 4.436780 seconds.
Compute the gradient of the solution wrt the parameters...
Elapsed time is 23.585154 seconds.
Compute the hessian of the solution wrt the parameters...
Elapsed time is 97.753342 seconds.
Solving C model ...
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000014 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000006 seconds.
Newton-Armijo solver
Iteration ||F(x_k)|| Steps
0 1.2193e-02 0
factorize Jacobian matrix ...
Elapsed time is 918.252195 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000033 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000009 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000015 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000007 seconds.
1 1.1695e-02 0
factorize Jacobian matrix ...
Elapsed time is 914.276151 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000041 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000011 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000015 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000006 seconds.
2 1.9138e-03 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000015 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000003 seconds.
3 1.3465e-03 0
factorize Jacobian matrix ...
Elapsed time is 897.176839 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000027 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000011 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000017 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000007 seconds.
4 5.9894e-04 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000013 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000006 seconds.
5 3.5496e-04 0
factorize Jacobian matrix ...
Elapsed time is 912.043570 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000035 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000010 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000016 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000006 seconds.
6 5.5135e-05 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000017 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000006 seconds.
7 1.7507e-05 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000010 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000003 seconds.
8 5.1336e-06 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000008 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000003 seconds.
9 1.3863e-06 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000011 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000003 seconds.
10 3.5546e-07 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000014 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000003 seconds.
11 8.8835e-08 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000011 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000002 seconds.
12 2.2076e-08 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000010 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000002 seconds.
13 5.5251e-09 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000016 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000006 seconds.
14 1.3994e-09 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000013 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000005 seconds.
15 3.5805e-10 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000017 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000004 seconds.
16 9.2164e-11 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000013 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000005 seconds.
17 2.3773e-11 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000013 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000004 seconds.
18 6.0893e-12 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000014 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000004 seconds.
19 1.5877e-12 0
Compute the gradient of the solution wrt parameters ...
Elapsed time is 0.000016 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 0.000005 seconds.
20 4.7245e-13 0
reset the global variable for the next call eqCycle.
factorize Jacobian matrix ...
Elapsed time is 863.131733 seconds.
Compute the gradient of the solution wrt parameters ...
Elapsed time is 184.140662 seconds.
Compute the hessian of the solution wrt the parameters ...
Elapsed time is 1363.570053 seconds.
Solving O model ...
Newton-Armijo solver
Iteration ||F(x_k)|| Steps
0 3.9132e-03 0
1 3.3859e-03 1
2 1.8399e-03 0
3 1.1473e-03 1
4 8.9825e-04 1
5 7.5800e-04 2
6 6.9120e-04 2
7 6.4705e-04 2
8 6.1991e-04 2
9 5.8800e-04 2
10 5.6486e-04 2
11 5.5868e-04 2
12 5.3384e-04 2
13 5.1417e-04 2
14 4.8599e-04 2
15 4.6316e-04 2
16 4.4701e-04 2
17 4.2052e-04 2
18 4.0394e-04 2
19 3.9327e-04 2
20 3.7435e-04 2
21 3.5951e-04 2
22 3.4437e-04 2
23 3.3220e-04 2
24 3.1809e-04 2
25 3.0658e-04 2
26 2.9275e-04 2
27 2.7859e-04 2
28 2.6459e-04 2
29 2.5105e-04 2
30 2.3833e-04 2
31 2.2835e-04 2
32 2.1677e-04 2
33 2.0579e-04 2
34 1.9647e-04 2
35 1.8732e-04 2
36 1.7795e-04 2
37 1.6905e-04 2
38 1.6063e-04 2
39 1.5257e-04 2
40 1.4492e-04 2
O2model did not converge.
Using the time stepping method for initial O2 ...
dt = 3600.000000 nsteps = 100
....................................................................................................
dt = 86400.000000 nsteps = 100
....................................................................................................
dt = 2628000.000000 nsteps = 100
....................................................................................................
dt = 31536000.000000 nsteps = 200
........................................................................................................................................................................................................
dt = 126144000.000000 nsteps = 200
........................................................................................................................................................................................................
dt = 315360000.000000 nsteps = 300
............................................................................................................................................................................................................................................................................................................
Elapsed time is 599.841313 seconds.
Used O2(:,:,:,250) for initial O2 field of nsnew ...
Newton-Armijo solver
Iteration ||F(x_k)|| Steps
0 2.2592e-03 0
1 2.2355e-03 0
2 1.3188e-03 1
3 9.4577e-04 1
4 7.7328e-04 2
5 6.9115e-04 2
6 6.6921e-04 2
7 6.1935e-04 2
8 5.9021e-04 2
9 5.7575e-04 2
10 5.5707e-04 2
11 5.3474e-04 2
12 5.1639e-04 2
13 4.8471e-04 2
14 4.6280e-04 2
15 4.4260e-04 2
16 4.1745e-04 2
17 4.0091e-04 2
18 3.8147e-04 2
19 3.6714e-04 2
20 3.4868e-04 2
21 3.3505e-04 2
22 3.1929e-04 2
23 3.0486e-04 2
24 2.9464e-04 2
25 2.8441e-04 2
26 2.7442e-04 2
27 2.6043e-04 2
28 2.4744e-04 2
29 2.3584e-04 2
30 2.2391e-04 2
31 2.1260e-04 2
32 2.0224e-04 2
33 1.9220e-04 2
34 1.8253e-04 2
35 1.7353e-04 2
36 1.6482e-04 2
37 1.5655e-04 2
38 1.4869e-04 2
39 1.3791e-04 2
40 1.2685e-04 2
Used O2(:,:,:,500) for initial O2 field of nsnew ...
Newton-Armijo solver
Iteration ||F(x_k)|| Steps
0 3.5050e-03 0
1 1.9476e-03 0
2 1.2651e-03 1
3 9.3798e-04 1
4 7.8671e-04 2
5 7.0825e-04 2
6 6.7379e-04 2
7 6.4210e-04 2
8 6.0551e-04 2
9 5.8483e-04 2
10 5.5254e-04 2
11 5.4513e-04 2
12 5.1025e-04 2
13 4.8790e-04 2
14 4.6729e-04 2
15 4.4317e-04 2
16 4.2681e-04 2
17 4.0225e-04 2
18 3.8569e-04 2
19 3.6707e-04 2
20 3.4910e-04 2
21 3.3664e-04 2
22 3.2939e-04 2
23 3.1318e-04 2
24 3.0365e-04 2
25 2.8906e-04 2
26 2.7498e-04 2
27 2.6127e-04 2
28 2.4876e-04 2
29 2.3619e-04 2
30 2.2423e-04 2
31 2.1292e-04 2
32 2.0206e-04 2
33 1.9187e-04 2
34 1.8222e-04 2
35 1.7309e-04 2
36 1.6154e-04 2
37 1.5005e-04 2
38 1.3399e-04 2
39 1.2443e-04 2
40 1.1430e-04 2
Used O2(:,:,:,750) for initial O2 field of nsnew ...
Newton-Armijo solver
Iteration ||F(x_k)|| Steps
0 4.7789e-03 0
1 4.5251e-03 0
2 2.0197e-03 0
3 1.0170e-03 1
4 8.5852e-04 1
5 7.3254e-04 2
6 6.7287e-04 2
7 6.4611e-04 2
8 6.0489e-04 2
9 5.7984e-04 2
10 5.5540e-04 2
11 5.4647e-04 2
12 5.0215e-04 2
13 4.9016e-04 3
14 4.8660e-04 2
15 4.6117e-04 2
16 4.4350e-04 2
17 4.2098e-04 2
18 4.0121e-04 2
19 3.8375e-04 2
20 3.7087e-04 2
21 3.5040e-04 2
22 3.3469e-04 2
23 3.2391e-04 2
24 3.1180e-04 2
25 2.9759e-04 2
26 2.8314e-04 2
27 2.8173e-04 3
28 2.8031e-04 3
29 2.7891e-04 3
30 2.7869e-04 4
31 2.7115e-04 2
32 2.5981e-04 2
33 2.5182e-04 2
34 2.3921e-04 2
35 2.2743e-04 2
36 2.2403e-04 3
37 2.1267e-04 2
38 2.0194e-04 2
39 1.9181e-04 2
40 1.8221e-04 2
Used O2(:,:,:,1000) for initial O2 field of nsnew ...
Newton-Armijo solver
Iteration ||F(x_k)|| Steps
0 4.7789e-03 0
1 4.5251e-03 0
2 2.0197e-03 0
3 1.0170e-03 1
4 8.5852e-04 1
5 7.3254e-04 2
6 6.7287e-04 2
7 6.4611e-04 2
8 6.0489e-04 2
9 5.7984e-04 2
10 5.5540e-04 2
11 5.4647e-04 2
12 5.0215e-04 2
13 4.9016e-04 3
14 4.8660e-04 2
15 4.6117e-04 2
16 4.4350e-04 2
17 4.2098e-04 2
18 4.0121e-04 2
19 3.8375e-04 2
20 3.7087e-04 2
21 3.5040e-04 2
22 3.3469e-04 2
23 3.2391e-04 2
24 3.1180e-04 2
25 2.9759e-04 2
26 2.8314e-04 2
27 2.8173e-04 3
28 2.8031e-04 3
29 2.7891e-04 3
30 2.7869e-04 4
31 2.7115e-04 2
32 2.5981e-04 2
33 2.5182e-04 2
34 2.3921e-04 2
35 2.2743e-04 2
36 2.2403e-04 3
37 2.1267e-04 2
38 2.0194e-04 2
39 1.9181e-04 2
40 1.8221e-04 2
Warning: O2model did not converge even after the time stepping method.
>>
Description
I am confused about how the
ResetParfunction affects thefminuncsearch algorithm. During an optimization where I switched theC2Pparameterization to be a global constant optimizable parameter, the optimization ended at what it considers to be a local minimum. However, the objective function is significantly larger than it was at the initial iteration. I am wondering if this might be related to theResetParfunction.Details
Below is a subset of the log file from the optimization run:
Notice how the
ResetParfunction repeatedly resets thebPandbP_Tparameters (pindx 4&5) in the first 9 iterations. For these iterations, the objective function value using the reset parameters is reasonably small. Then, from the 10th iteration onwards, the parameters do not get reset (becauseResetParis only executed inneglogpostif iter < 10). In iteration 10, the two parameters that had been reset previously make large jumps. (bP_Tjumps from 0.284 to -1.0 andbPjumps from 0.894 to 4.97).My first guess is that
fminuncmakes its next guess based on the values it initially suggested, rather than the reset values. When the objective function value decreases after resetting some of the parameter values,fminuncmight proceed along its initial search direction, thinking that the parameter values it tried moved the function closer to a minimum, despite the model not using those unrealistic parameters to calculate the objective function, gradient, and Hessian.Potential Solutions
If
ResetParis a problem forfminunc, we might want to consider removingResetParto avoid this issue. Or, if there are potentially bad values that could cause the model to crash, another option would be to impose a very large objective function value whenResetParfinds bad values, instead of running the model with reset values.