Corner case of integer variables with boundaries

[nikohansen]

Related to #259.

For an upper variable bound at 1.5, an infeasible value of 1.6 is repaired to 1.5 which is then rounded to 2.0 (which is not an issue with BoundTransform.repair). If 1 is the optimum, this means that the optimum is not evaluated anymore (for unjustifiable reasons) unless the mean gets feasible again.

This is also a potential issue on the bbob-mixint test suite (with BoundPenalty handling).

A possible fix is to amend the bounds for integer variables (phenotype indices) by amending the bounds option.

[nikohansen]

In the integer_centering module the corner case is addressed like

        idx = np.mod(bounds, 1) == 0.5
        bounds[idx] += (1 if which == 'lower' else
            -1) * self.params['bounds_offset'] * np.maximum(1, np.abs(bounds[idx]))

[nikohansen]

Fixed with commit 8b11737.