Multi-Objective Optimization¶
Often we do not only want to optimize just a single objective, but multiple instead. SMAC offers a multi-objective optimization interface to do exactly that. Right now, the algorithm used for this is a mean aggregation strategy or ParEGO [Know06]. In both cases, multiple objectives are aggregated into a single scalar objective, which is then optimized by SMAC. However, the run history still keeps the original objectives.
The basic recipe is as follows:
Specify the objectives in the scenario object as list. For example,
Scenario(objectives=["obj1", "obj2"]).Make sure that your target function returns a cost dictionary containing the objective names as keys and the objective values as values, e.g.
{'obj1': 0.3, 'obj2': 200}. Alternatively, you can simply return a list, e.g.,[0.3, 200].Now you can optionally pass a custom multi-objective algorithm class to the SMAC facade (via
multi_objective_algorithm). In all facades, a mean aggregation strategy is used as the multi-objective algorithm default.
Warning
The multi-objective algorithm influences which configurations are sampled next. More specifically, since only one surrogate model is trained, multiple objectives have to be scalarized into a single objective. This scalarized value is used to train the surrogate model, which is used by the acquisition function/maximizer to sample the next configurations.
You receive the incumbents (points on the Pareto front) after the optimization process directly. Alternatively, you can
use the method get_incumbents in the intensifier.
smac = ...
incumbents = smac.optimize()
# Or you use the intensifier
incumbents = smac.intensifier.get_incumbents()
We show an example of how to use multi-objective with plots in our examples.