[GENERAL SUPPORT]: Reference Point for Multi-Objective Bayesian Optimization about ax

[mgrange1998]

Hi, thank you for posting the question.

The main impact of selecting the reference point is how candidates are generated over the cross of the experiment run. See this wiki section for an example of how tighter objective thresholds affect the search space- mainly a tighter concentration of points in the valid region.

Additionally, the difference in reference point affects which points are included in the Pareto frontier- looser constraints means more points on the frontier.

Usually the reference point is within the bound of the search space, and is more restrictive than the outcome constraints, so a typical example would be something like A=15, B=2.25.

[renaissanceM4n]

Thank you for your feedback @mgrange1998 . An addition to the last statement. I am not restricting my metrics but using an artificial restriction of my objectives A and B, so I must correct myself.
Unfortunately, I am generating very unusual Pareto fronts. Perhaps the restrictions are too drastic and the target values are completely contradictory? The fronts display points that are far outside the range defined by the reference points. I can't quite explain this at the moment. In the Pareto front shown, the reference point for the objective on the y-scale is set at 1 and yet points far above this are still displayed as Pareto-optimal points.
I have modified the objective on the y-scale during optimisation, which resulted in the Pareto front shown here, so that the evaluation function only ever returns the deviation from the value 0. In this way, I want to generate optimal parameterisations for the output value 0 of this metric. Could it also be a problem that the surrogate model shows me negative values, although my return value cannot be less than zero?
Or are there any recommendations for obtaining the optimisation result described?