search

AequilibraE / aequilibrae

aequilibrae - Python package for transportation modeling
https://www.aequilibrae.com
Other
166 stars 38 forks source link

Select link cwf bfw bug #500

Closed Jake-Moss closed 8 months ago

Jake-Moss commented 8 months ago

Addressing the bug report raised in https://github.com/AequilibraE/aequilibrae/issues/493

Current thoughts are that the blending of the previous and current iteration is failing. The fix in 4894f619ff2393ad0244d44db8be8343d57748e5 doesn't seem to be enough, though it does extend the number of iterations where the results agree from 1 to 4.

Jake-Moss commented 8 months ago

Overall the root of the issue was that the select link wasn't taking into account the previous step directions in it's calculations. In principle, the patch that modified the select to do just that should have worked, however, the matrix that stored the previous step directions was allocated on the wrong object and, while the object lived the entire duration of the assignment, the matrix was being zeroed at each iteration. I'd overlooked this as it did improve the stability up to 4 iterations, but Jan pointed out to me that the beta's being used were 1.0 (and 0.0 respectively) up until the 5th iteration where the step direction truly started to matter.

Another two bugs I found are patched.

  1. The aon object was being leaked outside of the traffic class for loop, this meant that the last traffic class was being used for all select links analysis'. The aon objects are now stored and reused, this may increase memory usage during the assignment as GC can no longer clean the memory allocated by these objects up at the end of every AON assignment, instead it will at the end of all assignments.

  2. As noted in the comment pre_previous_step_direction was being used in place of previous_step_direction.