- Define a multi-output function as
$f: \mathbb{R}^D \rightarrow \mathbb{R}^M$ , - Assume we run
$N$ independent optimization runs and obtain the Pareto sets for each run$\forall i \in \{1,\dots,M\}, \mathcal{F}_i \subseteq \mathbb{R}^M$ , - Define the objective vector
$\boldsymbol{f} \in \mathbb{R}^M$ weakly dominates a vector$\boldsymbol{y}$ in the objective space if and only if$\forall m \in \{1,\dots,M\}, f_m \leq y_m$ and notate it as$\boldsymbol{f} \preceq \boldsymbol{y}$ , and - Define a set of objective vectors
$F$ weakly dominates a vector$\boldsymbol{y}$ in the objective space if and only if$\exists \boldsymbol{f} \in F, \boldsymbol{f} \leq \boldsymbol{y}$ and notate it as$F \preceq \boldsymbol{y}$
As seen in the figure below, the attainment surface is the surface in the objective space that we can obtain by splitting the objective space like a step function by the Pareto front solutions yielded during the optimization.
It is simple to obtain the attainment surface if we have only one experiment;
however, it is hard to show the aggregated results from multiple experiments.
To address this issue, we use the
Credit: Figure 4. in Indicator-Based Evolutionary Algorithm with Hypervolume Approximation by Achievement Scalarizing Functions.
First, we define the following empirical attainment function:
The
Note that as we only have
The best, median, and worst attainment surfaces could be fetched by
Please check the following references for more details:
[2] An Approach to Visualizing the 3D Empirical Attainment Function