Questions about multi-objective optimization.

[showgood163]

Hi, newbie here!

I have a question about Multi-objective optimization mentioned here.

Does Ax have the ability to simultaneously optimize more than one objectives? If so, how do I achieve that? Is there any suggestions/tutorials?

[BCJuan]

Hi,

There is an objetive function class named ScalarizedObjective in ax.core.objective which allows to establish several metrics and the weights of each.

https://ax.dev/api/core.html?highlight=scalarized%20objective#ax.core.objective.ScalarizedObjective

Indeed, you could also change the weights of each of the objectives whenever you desired to by accessing them from the experiment configuration in the squence generator or in the experiment directly. Take a look to the "Human in the loop tutorial" to see the change in action.

In the tutorial:

constraint_5 = OutcomeConstraint(metric=constraint_metric, op=ComparisonOp.LEQ, bound=5)
constraint_5_results = gp.gen(
    n=3, 
    optimization_config=OptimizationConfig(
        objective=Objective(objective_metric),
        outcome_constraints=[constraint_5]
    )
)

where they change the objective for the generator of new points to evaluate.

[sdsingh]

Hey there! Thanks for the answer, @BCJuan!

Typically, our users deal with multiple-objective optimization by specifying both an Objective and a series of OutcomeConstraints. The "Human in the loop tutorial" demonstrates how one can set and adjust OutcomeConstraints/Objectives to obtain a different set of results.

In that tutorial, we don't actually use ScalarizedObjective. Typically, users find it difficult to come up with a single scalarization of their multi-objective problem, and thus prefer the Objective + OutcomeConstraint formulation.

If you would prefer to use ScalarizedObjective, you could attempt to change the weights within the objective, and repeatedly generate candidates, in a similar fashion to the tutorial.