[Bug report] Optimal solution is slightly infeasible

[paulduf]

Describe the bug

The optima solution obtained with the (hidden) method problem._best_parameter for problem f18 in 20-D has some very small positive value for some constraints. This prevents the triggering of the final target (and maybe some others) when evaluating what is supposed to be a reference for the global optimum.

Your System (please complete the following information when it seems relevant):

• Operating system: Ubuntu 18.04
• COCO version: 2.6.2

Minimal Reproducible Example

import cocoex
import numpy as np
suite = cocoex.Suite("bbob-constrained", "", "")
problem = suite.get_problem_by_function_dimension_instance(18, 20, 1)
problem._best_parameter('print')
_xopt = np.loadtxt('._bbob_problem_best_parameter.txt')
problem(_xopt)
problem.final_target_hit
max(problem.constraint(_xopt))

Outputs:

0
1.5969018398508795e-13

[nikohansen]

It looks like subtracting 1e-12 from the constraint values solves the issue on all (first) 15 function instances in all dimensions. This should have practically no influence on the performance evaluation. In the construction, constraints are multiplied by up to about 1e5.
EDIT: it should be 1e-11 and I am not sure why the multiplier is relevant.

[paulduf]

Related plot: CDF of max(g-values) over the first 100 instances of all problems in all dimensions (on a symlog scale).

[nikohansen]

I thought changing the 0.0 in this line https://github.com/numbbo/coco/blob/master/code-experiments/src/c_linear.c#L75 to -1e-11 should do the trick, but it seems like this code has no effect at all.

[nikohansen]

I implemented a different solution 9294678 which does in each coordinate this transformation

and returns exactly zero in the above example. The distribution of the maximal constraint in the optimum over all instances looks like

and the maximal value is zero.

A condition of x[i] > 1e-12 instead of > 1e-11 gives a maximum constraint violation of 6.83612196645526e-12.

[nikohansen]

@paulduf, @brockho how do we check that 9294678 doesn't break the regression test for the testbed?

[brockho]

python do.py test-suites should do what you want, shouldn't it? The documentation says "Runs regression test on all benchmark suites".

[nikohansen]

The regression test passes. It tests like np.abs(x - y) < 1e-9, so 1e-11 should not matter even when we test this very position.