This directory contains files and results for experiments using arcopt on Michael Haythorpe’s HCP objective function.
The experimental results are stored in sub dirs of results. This file contains some short notes describing the experiments.
All graphs supplied by Haythorpe and indicated as having a ham. cycle were run with both arcopt and ipopt.
Individual graphs were given to hcp_arcopt.m and hcp_ipopt.m. The drivers for the experiment are drv_arcopt.m and drv_ipopt.m.
The resulting tables and summary were produced by drv_results.m
This experiment did two things:
- printed output for each solve of hcp_data with arcopt
- looked at the overall results for 10, 12, and 14 node graphs
MH sent 4 new graphs:
- graph24.m
- graph30.m
- graph30degree4.m
- graph38.m
I ran arcopt and ipopt on these graphs with x0:
- as the analytic center of the feasible region
- as a random feasible point generated with rand and refined by solving min ||x0-v|| s.t. A*x0 == b, 0 <= x0 <= 1 with cvx
The result was fewer function evals and comparable results. It only found an HC on 8/20 tries for the 38 node graph as opposed to 10/20 with function evals always on.
The code simple plots data from the table:
number of nodes | degree | nvar | fev |
---|---|---|---|
10 | 3 | 30 | 16 |
12 | 3 | 36 | 19 |
14 | 3 | 42 | 23 |
24 | 3 | 72 | 34 |
30 | 3 | 90 | 39 |
30 | 4 | 120 | 71 |
38 | 3 | 114 | 48 |
This test attempts to solve the problem with only the stochastic constraint.
I had to change hcp_con to make this work.
I copied data from exp on 2012-01-18.