Introduction
In multi-objective optimization, the result is typically a Pareto front—a set of solutions representing optimal trade-offs where no single objective can be improved without degrading another. While these solutions are theoretically equivalent in terms of optimality, their distribution in the objective space is often highly irregular in real-world applications.
In fields such as medical imaging, solutions frequently "clump" together. A large number of solutions might be concentrated in a small region, while other areas remain sparse. This irregularity makes it difficult for human experts to navigate the solution space. For instance, using a slider to select a solution might result in skipping over hundreds of similar options with a tiny movement, or seeing no change at all over a large movement. This paper proposes a method to redistribute these points onto a uniform simplex to facilitate intuitive selection.
Methodology
The authors transform the visualization task into a multi-objective optimization problem itself. The goal is to map an $m$-dimensional Pareto front onto an $(m-1)$-simplex (e.g., a 3D front onto a 2D triangle) while balancing two conflicting objectives:
- Uniformity ($F_{spread}$): Points should be spread as evenly as possible across the simplex. This is modeled as a relaxed version of the "Circles-In-A-Square" packing problem, using a $p$-norm (with $p=4$) to ensure a smooth fitness landscape for the solver.
- Topology Preservation ($F_{stress}$): The relative neighborhood structure of the original space must be maintained. Solutions that were close in the high-dimensional space should remain close in the 2D visualization. This is measured using a modified Sammon stress metric.
The optimization is performed using MO-RV-GOMEA, a state-of-the-art evolutionary algorithm. To improve efficiency, the population is initialized using barycentric weights derived from the original coordinates.
Results and Discussion
The method was evaluated on benchmark problems from Deformable Image Registration (DIR) involving 3D breast MRI scans. The objectives included image dissimilarity, deformation magnitude, and guidance error.
As shown in the comparison above, a naive projection (Figure 2) preserves the clumping of the original space, making many solutions indistinguishable. In contrast, the proposed redistribution (Figure 3) creates a near-uniform layout. Color-coding based on original objective values confirmed that the local neighborhood structure was preserved, with extreme solutions correctly positioned at the vertices of the triangle.
This uniform distribution allows for "intuitive navigation." Every part of the visualization becomes equally accessible and distinct, enabling the use of standard UI elements like barycentric sliders where visual distance correlates linearly with the perceived change in the solution.
Conclusion
By treating visualization as an optimization problem, this research provides a way to make complex Pareto fronts navigable for human experts. While the current approach introduces some computational overhead and requires manual selection from a "Pareto front of visualizations," it significantly improves the usability of multi-objective optimization results in practical domains. Future work may explore discrete permutation alternatives to further enhance efficiency.
Summary updated