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:
Make sure that your target function returns a cost dictionary containing the objective names as keys and the objective values as values, e.g.
{'myobj1': 0.3, 'myobj2': 200}. Alternatively, you can simply return a list, e.g.,[0.3, 200].When instantiating SMAC pass the names of your objectives to the scenario object via the
objectivesargument, e.g.,multi_objectives: ["myobj1", "myobj2"].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
Depending on the multi-objective algorithm, the incumbent might be ambiguous because there might be multiple
incumbents on the Pareto front. Let’s take ParEGO for example:
Everytime a new configuration is sampled, the weights are updated (see runhistory encoder). Therefore, calling
the get_incumbent method in the runhistory might return a different configuration based on the internal state
of the multi-objective algorithm.
You can use get_pareto_front in the run history to get the configurations on the Pareto front.
smac = ...
smac.optimize()
for incumbent, costs in smac.get_pareto_front():
print(incumbent, costs)
We show an example of how to use multi-objective with plots in our examples.