Disable named expressions for MB+LD2 - #71
Conversation
Runtime resultsThese are MB+LD2 With named expressionsIdentifier ncalls cumtime percall %
-----------------------------------------------------------------------------------------------------
root 1 1.288 1.288 100.0
------------------------------------------------------------------------------------------------
mb-steady-linear-d2 1 1.288 1.288 100.0
-----------------------------------------------------------------------------
elimination 1 1.155 1.155 89.7
------------------------------------------------------------------
eliminate 13 0.182 0.014 15.7
eliminate-nodes 13 0.133 0.010 11.5
get-d1 14 0.492 0.035 42.6
get-d2 2 0.107 0.053 9.2
igraph 1 0.098 0.098 8.5
linear-igraph 1 0.103 0.103 8.9
subgraph 1 0.040 0.040 3.4
other n/a 0.002 n/a 0.1
==================================================================
solver 1 0.084 0.084 6.5
------------------------------------------------------------------
solve 1 0.012 0.012 14.2
---------------------------------------------------
constraint-jac 10 0.000 0.000 3.9
constraints 9 0.001 0.000 6.5
intermediate-cb 9 0.000 0.000 0.1
lagrangian-hess 8 0.002 0.000 14.7
objective 9 0.000 0.000 0.6
objective-grad 10 0.000 0.000 1.1
other n/a 0.009 n/a 73.1
===================================================
other n/a 0.072 n/a 85.8
==================================================================
other n/a 0.049 n/a 3.8
=============================================================================
other n/a 0.000 n/a 0.0
================================================================================================
=====================================================================================================Without named expressionsIdentifier ncalls cumtime percall %
-----------------------------------------------------------------------------------------------------
root 1 1.296 1.296 100.0
------------------------------------------------------------------------------------------------
mb-steady-linear-d2 1 1.296 1.296 100.0
-----------------------------------------------------------------------------
elimination 1 1.161 1.161 89.6
------------------------------------------------------------------
eliminate 13 0.178 0.014 15.4
eliminate-nodes 13 0.118 0.009 10.2
get-d1 14 0.490 0.035 42.2
get-d2 2 0.125 0.062 10.7
igraph 1 0.100 0.100 8.6
linear-igraph 1 0.108 0.108 9.3
subgraph 1 0.041 0.041 3.5
other n/a 0.002 n/a 0.1
==================================================================
solver 1 0.088 0.088 6.8
------------------------------------------------------------------
solve 1 0.011 0.011 13.0
---------------------------------------------------
constraint-jac 10 0.000 0.000 4.0
constraints 9 0.001 0.000 7.1
intermediate-cb 9 0.000 0.000 0.1
lagrangian-hess 8 0.002 0.000 15.4
objective 9 0.000 0.000 0.7
objective-grad 10 0.000 0.000 1.1
other n/a 0.008 n/a 71.6
===================================================
other n/a 0.076 n/a 87.0
==================================================================
other n/a 0.048 n/a 3.7
=============================================================================
other n/a 0.000 n/a 0.0
================================================================================================
=====================================================================================================No appreciable difference. |
Convergence resultsAll MB results, run on Darwin. With named expressions for method (main branch) model method n-success n-total percent-success ave-elim-time ave-solve-time
0 mb-steady no-elim 88 121 72.727273 0.000039 1.056338
1 mb-steady d1 89 121 73.553719 4.647681 1.095656
2 mb-steady ecd2 89 121 73.553719 4.475969 1.124023
3 mb-steady linear-d2 88 121 72.727273 5.194716 0.998081
4 mb-steady d2 96 121 79.338843 NaN 0.918361
5 mb-steady greedy 105 121 86.776860 1.365976 0.845214
6 mb-steady matching 101 121 83.471074 2.431153 1.042989Without named expressions for LD2 (this branch) model method n-success n-total percent-success ave-elim-time ave-solve-time
0 mb-steady no-elim 88 121 72.727273 0.000031 1.002411
1 mb-steady d1 89 121 73.553719 4.452170 1.033757
2 mb-steady ecd2 89 121 73.553719 4.469480 0.948522
3 mb-steady linear-d2 90 121 74.380165 5.077071 0.925272
4 mb-steady d2 96 121 79.338843 NaN 0.744996
5 mb-steady greedy 105 121 86.776860 1.353755 0.876362
6 mb-steady matching 101 121 83.471074 2.430085 0.904619Without named expressions, we LD2 converges two extra instances. |
|
In WARNING: pipeline, no-elim: success=True but feasible=False at index 79; infeasibility=0.000100078221294
WARNING: pipeline, ecd2: success=True but feasible=False at index 79; infeasibility=0.000100112039945
WARNING: pipeline, d2: success=True but feasible=False at index 79; infeasibility=0.0001000009942799I think I'm fine with accepting these instances. These are likely just "acceptable" infeasibilities (to IPOPT) that get exacerbated by unscaling the model. |
WARNING: pipeline, no-elim: success=True but feasible=False at index 79; infeasibility=0.000100078221294
WARNING: pipeline, ecd2: success=True but feasible=False at index 79; infeasibility=0.000100112039945
WARNING: pipeline, d2: success=True but feasible=False at index 79; infeasibility=0.0001000009942799These infeasibilities seem to be due to bound violations of flow rates. The variables have values around 10000 and the violations are of about |
|
This is waiting on a re-run of the convergence results with |
|
Thankyou Robby for debugging this and rerunning the results! The relative tolerance check sounds logical. Looking forward to seeing the results. |
| vars_violating_bounds.append((var, var.lb, lb_diff)) | ||
| relative_lb_diff = lb_diff / abs(var.lb) if var.lb != 0 else lb_diff | ||
| if lb_diff < - tolerance and relative_lb_diff < - tolerance: | ||
| vars_violating_bounds.append((var, var.lb, min(lb_diff, relative_lb_diff))) |
There was a problem hiding this comment.
In the warning message, maybe its worth printing that the infeasibility check failed in the absolute or relative sense.
There was a problem hiding this comment.
In this function, we only print the vague message "there are variables violating their bounds." I don't think we ever print a variable-specific warning message, so I don't see a good place to distinguish.
|
The other warnings with WARNING: pipeline, no-elim: success=True but feasible=False at index 103; infeasibility=6.353730395858292e-05 [681/9344]
WARNING: pipeline, ecd2: success=True but feasible=False at index 82; infeasibility=2.901628613471985e-05
WARNING: pipeline, linear-d2: success=True but feasible=False at index 81; infeasibility=1.3452023267745972e-05
WARNING: pipeline, d2: success=True but feasible=False at index 112; infeasibility=2.026371657848358e-05
WARNING: pipeline, greedy: success=True but feasible=False at index 2; infeasibility=9.983719792217016e-05
WARNING: pipeline, matching: success=True but feasible=False at index 2; infeasibility=9.98677860479802e-05For no-elim, warnings are due to material balance constraints where we equate two terms of magnitude ~1000. For matching, violations are due to bounds on an eliminated variable (now an inequality constraint) and would likely disappear with a relative tolerance check on this constraint as well. I'm re-running the results, but I'm leaning towards just setting |
|
We still have the following warnings after adding a relative tolerance check for inequality constraints. WARNING: pipeline, no-elim: success=True but feasible=False at index 103; infeasibility=6.353730395858292e-05
WARNING: pipeline, ecd2: success=True but feasible=False at index 82; infeasibility=2.901628613471985e-05
WARNING: pipeline, linear-d2: success=True but feasible=False at index 81; infeasibility=1.3452023267745972e-05
WARNING: pipeline, d2: success=True but feasible=False at index 112; infeasibility=2.026371657848358e-05 |
We could probably detect sum expressions in equality constraints with RHS=0 and do a relative tolerance check between the first two terms. I think this would catch these material balance discrepancies, at least. I'm going to merge this PR as-is, but I may add this check in another PR. |
|
So we don't have any warnings with relative tolerance implemented and setting feastol=1e-4? |
|
Yes. All violations are now less than |
2 participants
There is some odd behavior for the MB+LD2 model-method combination. Here are a few lines of output from
analyze_solvetime.py:The first line above is the last line of output from IPOPT: We converge to an optimal solution, but our post-solve constraint evaluation identifies an infeasible solution. What's unusual is that, on further inspection, the constraint violation magnitude is 0.17. This is much too large to be a difference due to scaled/unscaled models, IPOPT's bound relaxation, or any other solver-specific idiosyncrasy. I've found that this behavior only occurs when we use named expressions, so I'm patching our scripts to avoid using named expressions for this model-method combination.
I've looked into this discrepancy a bit and I think this is most likely an NL writer bug. I will open a Pyomo issue once I'm happy with this workaround.
This needs: