# Visualizing the diversity of a Pareto front approximation using RadVis

During many-objective search we seek to find approximate Pareto fronts that are convergent, meaning they are close to the “true” Pareto front, and diverse, meaning they contain solutions that they uniformly cover the full range of the “true” front. This post will focus on examining the latter criteria for high dimensional problems using Radial Visualization (RadVis; Hoffman et al., 1997), a visualization technique that transforms a solution set into radial coordinates. I’ll discuss the theory behind RadVis, provide some example code, and discuss potential applications, limitations and extensions.

### Background

When performing runtime diagnostics (a subject of two of my recent posts, which you can find here and here), we often asses MOEA performance using hypervolume because it captures both the convergence and diversity of the approximate Pareto front. Though tracking hypervolume provides us with an aggregate measure of diversity, it does not provide information on how the solutions are spread across the objective space. For low dimensional problems (two or three objectives), the diversity of the approximation set can easily be discerned through visualization (such as 2-D or 3-D scatter plots). For many-objective problems however, this becomes challenging. Tools such as parallel coordinate plots, bubble plots and glyph plots can allow us to visualize the solutions within an approximation set in their full dimensionality, but they are limited in their ability to convey information on diversity in high dimensional space. An alternative strategy for examining diversity within a solution set is to transform the objective space into a form coordinate system that highlights diversity. One common transformation is to view the set using radial coordinates.

RadVis uses an analogy from physics to plot many dimensional data. Each of n-objectives are plotted uniformly around the circumference of a unit circle while, each solution is represented as a point that is attached to all objectives by a set of n hypothetical springs (one spring is attached to each objective-solution pair). A solution’s objective value for the ith objective defines the spring constant exerted by the ith spring. A point’s location thus represents the equilibrium point between all springs (Hoffman et al., 1997). For example, a point that has equal objective values for all n-objectives will lie directly in the center of the plot, while a point that maximizes one objective and minimizes all others will lie on the circumference of the circle at the location specified for the maximized objective.

Below, I’ve plotted RadVis plots for a 5 objective instance of DTLZ2 at runtime snapshots from 1,000 function evaluations (NFE), 5,000 NFE, 10,000 NFE and 50,000 NFE. Notice how the approximate front for 1,000 NFE is sparse and skewed toward one side, while the front for 50,000 NFE is fairly uniform and centered. DTLZ2 is a test problem with a known true Pareto front which is continuous and uniformly distributed, using this information, we can see that the algorithm has likely improved its approximation set over the course of the run.

I coded this example using the Yellowbrick machine learning visualization toolkit from sklearn. Code for this example has been added to my runtime diagnostics Github repository, which can be found here.

### Caution! RadVis provides no information on convergence

It’s important to note that RadVis provides no information about solution convergence since the location of each point is dependent on equilibrium between all objectives. A set of solutions that performs poorly across all objectives will look similar to a set of solutions that performs well across all objectives. For this reason, RadVis should never be used alone to assess the quality of a Pareto front approximation. To include convergence information in RadVis, Ibrahim et al,. (2016) and Ibrahim et al., 2018 have proposed 3d-RadVis and 3D-RadVis antenna, extensions of RadVis methodology which augment Radvis by providing convergence information to decision makers. These methodologies may be subject to future posts.

References

Hoffman, P., Grinstein, G., Marx, K., Grosse, I., & Stanley, E. (1997, October). DNA visual and analytic data mining. In Proceedings. Visualization’97 (Cat. No. 97CB36155) (pp. 437-441). IEEE.

Ibrahim, A., Rahnamayan, S., Martin, M. V., & Deb, K. (2016, July). 3D-RadVis: Visualization of Pareto front in many-objective optimization. In 2016 IEEE Congress on Evolutionary Computation (CEC) (pp. 736-745). IEEE.

Ibrahim, A., Rahnamayan, S., Martin, M. V., & Deb, K. (2018). 3D-RadVis Antenna: Visualization and performance measure for many-objective optimization. Swarm and evolutionary computation, 39, 157-176.

# MOEA Performance Metrics

“How well is my search algorithm doing?”, you ask.

Well, that’s a great question!

Let’s first look at the example of a single-objective, analytical formulation such as an LP.  Generally speaking we know there is either no solution, exactly one solution, or infinitely many solutions to the LP.  A solution is optimal, but it’s only optimal with respect how you defined your problem.  We’ll take the example of trying to optimize the production of items in a factory and show two potential issues with the classical approach.  First, an LP formulation of such a problem can’t consider economies of scale (i.e., the first unit costs just as much as the 100,000th unit).  You may need a more complicated model of your process, such as a full scale simulation model, to characterize your solution better.  Second, there may be multiple objectives to consider.  You can minimize cost in an LP, but what about resource usage? Time of production? And so on.

The approach we talk about here seeks to address these concerns.  Our approach is a simulation-optimization approach, which means the complex simulation model is embedded in the search — how the system is characterized in the simulation is included when you’re generating alternatives.  And it is many-objective, which means you can generate the explicit tradeoffs between conflicting objectives.  Such a tradeoff is called the Pareto-optimal frontier, and if you can’t prove you have the true Pareto optimal frontier you have the Pareto-approximate frontier.

But there is a challenge!  And the challenge relates to our original question.  For a difficult problem, it’s very hard to prove that you have the true Pareto optimal frontier.  Even if you have a best-known approximation to the frontier, how do you know “how good”  any single run of a MOEA has done, at getting you near that best-known approximation?

The best known set (black), and one approximation of it (gray). Preferred solutions are found toward the lower left corner (i.e., a min-min problem)

In the image above, we want to minimize both objectives.  So preferred solutions are near the lower left corner of the figure.  The best known approximation set is shown in black.  A given run of an algorithm produces the gray points.  There’s two concepts we can consider.  The first is convergence — how “close” did we come to finding the black set?  The second is diversity — are there points that are fully spread across the set?  Or is there only a cluster at, say, high values of the first objective?

Visual evaluation of your approximation set can start to help you answer these questions.  It’s especially helpful if you can visualize the MOEA’s performance during the search process itself.  But this can be challenging for high-dimensional problems.  Luckily there are quantitative metrics that can give you a feeling for both convergence and diversity of your approximation sets.

## Quantitative Metrics

For the rest of this post, I wanted to just give an informal description of three important performance metrics.  For a full description, please see a reference such as our recent comparison of MOEA performance on water problems, published in Advances in Water Resources.

Also, users of MOEAframework can easily calculate these metrics for their own problems!

### Generational Distance

Generational Distance illustration. Red lines show distance between approximation and the best approximation.

Generational Distance gives a good first approximation of the performance of your approximation set.  The red lines in the figure show a calculation of distance.  For each of the gray points, find the closest point in the best approximation.  The metric is then the average of these distances.  We consider this an “easy” metric to meet, because in a situation where you only have one solution close to the set, you’d have almost perfect Generational Distance performance.

### Epsilon Indicator

Epsilon Indicator is the worst-case distance that you’d have to translate your set to have it “dominate” the best approximation.

Epsilon indicator is considered a “harder” metric to meet.  It’s based on the idea of one set “dominating” another: having better performance in both objectives.  Near the top of the figure, we see that the gray points are close to the black points.  But on the lower right side of the figure, the gray points are much further from the black ones.  Epsilon indicator uses the worst-case distance that you would have to translate the approximation set, in order to have it dominate the best-known approximation (see the red arrow in the illustration).

The metric is harder to meet, because now if you have a large gap in your approximation, you will have poorer metric performance.

### Hypervolume

The hypervolume compares a multidimensional volume determined by the approximation (red) to the volume determined by the best known approximation (black), relative to a reference point.

The hypervolume metric captures both convergence and diversity.  It does this in a clever way — by looking at the multidimensional “volume” created by each set, relative to a reference point.  Such volumes are illustrated by the red and the black spaces on the figure.  You can see that you can’t really capture good hypervolume unless you have almost every point, across the full spread of objective function performance.  One downside is that the hypervolume is computationally intensive, but this is improving with new algorithms, some of which are supported by MOEAframework.