Crowding Distance Truncation in NSGA-II

• Crowding distance-based truncation is a diversity-preserving method in NSGA-II that quantifies candidate isolation across multiple objectives to ensure well-spread solutions.
• It leverages boundary assignment and neighbor-based differences to rank solutions, with improved variants like truthful crowding distance addressing clustering pitfalls.
• Empirical and theoretical analyses demonstrate that tailored truncation strategies can enhance Pareto front coverage and maintain superior diversity compared to uniform selections.

Crowding distance–based truncation is a selection and diversity-maintaining mechanism central to the Non-dominated Sorting Genetic Algorithm II (NSGA-II) family of multi-objective evolutionary algorithms. It quantifies the isolation of a candidate solution in objective space to guide which solutions survive to the next generation when the number of non-dominated solutions exceeds available slots. Crowding distance–based truncation aims to balance convergence toward the Pareto front with maintenance of diverse, well-spread solutions, and has led to a sequence of refinements and theoretical analysis, including improved definitions and provably sound variations for many-objective settings (Chu et al., 2018, Zheng et al., 2024, Ishibuchi et al., 24 Apr 2025).

1. Formal Definition of Crowding Distance–Based Truncation

Let $F$ be a non-dominated front of size $|F|$ under $m$ objectives. For each individual $j=1,\dots,|F|$ in $F$, denote its value on the $k$-th objective as $f_j^k$, with $f_{\min}^k$ and $f_{\max}^k$ the minimum and maximum values of $f^k$ over $F$.

Original NSGA-II Crowding Distance:

• For each objective $k$, sort $F$ by $f^k$ in ascending order.
• Set boundary points (minimum and maximum in each objective) to have $d_j = +\infty$.
• For each interior point $j$ (with neighbors $j-1$ and $j+1$), update:

$$d_j \leftarrow d_j + \frac{f^k_{j+1} - f^k_{j-1}}{f^k_{\max} - f^k_{\min}}$$

• The total crowding distance is the sum over all objectives.

Improved Crowding Distance (Chu et al., 2018):

• Replace the symmetric span by a forward difference:

$$d_j^{\mathrm{imp}} \leftarrow d_j^{\mathrm{imp}} + \frac{f^k_{j+1} - f^k_j}{f^k_{\max} - f^k_{\min}}$$

This biases the distance in favor of solutions closer to the Pareto front.

Truthful Crowding Distance (tCD) (Zheng et al., 2024):

• For each objective $i$, sort $S$ in descending order of $f_i$.
• For $j>1$, define the normalized $L_1$ distance between $S_{i.k}$ and $S_{i.j}$ as

$$d(S_{i.k}, S_{i.j}) = \sum_{a=1}^m \frac{|f_a(S_{i.k}) - f_a(S_{i.j})|}{f_a(S_{a.1})-f_a(S_{a.N})}$$

• The per-objective tCD is the minimum over all earlier $k<j$ in the sorted list.
• Final tCD for $x$ is $+\infty$ if $x$ is a boundary solution for any objective; otherwise, tCD$(x) = \sum_{i=1}^m \mathrm{tCD}_i(x)$.

2. Algorithmic Structure of Truncation in NSGA-II

The truncation operator is invoked when the non-dominated fronts $F_1, F_2, \dots$ exceed the intended population size $N$. The next generation is filled by:

• Sequentially adding entire fronts until reaching a front $F_j$ that would overflow $N$.
• Computing and assigning crowding distances $d_j$ to all members of $F_j$.
• Sorting $F_j$ in descending $d_j$ (with $+\infty$ first).
• Choosing the $N - |\bigcup_{l=1}^{j-1} F_l|$ solutions from the top of this order.

Boundary solutions (extremes in any objective) are assigned $d=+\infty$, and tie-breaking among equal $d_j$ is typically random or by stable sort. The time complexity remains $O(mN \log N)$.

When using tCD, the only change is substitution of the crowding distance computation subroutine, with otherwise unchanged selection and sorting logic (Zheng et al., 2024).

3. Theoretical Properties and Optimality Criteria

On linear Pareto fronts for two-objective problems, the crowding distance for an interior point $x_i$ is explicitly $d_i = 2(x_{i+1} - x_{i-1})$ (Ishibuchi et al., 24 Apr 2025). The optimization problem thus becomes maximizing the minimum three-point spacing.

Theoretical analysis demonstrates that the uniform spacing $x_i = (i-1)/(\mu-1)$, while intuitive, does not maximize the minimum crowding distance. The true optimum corresponds to clustered overlap: with $\mu$ solutions, the best configuration arranges them in $\lceil \mu/2 \rceil$ clusters equally distributed along the front, with (for even $\mu$) two solutions per cluster. The minimum crowding distance achieved in this way is

$\Delta^* = \frac{2}{\lceil \mu/2 \rceil - 1}$

In contrast, the uniform distribution gives $\Delta_{\mathrm{uni}} = 4/(\mu-1)$, and analytic and empirical results confirm $\Delta^* > \Delta_{\mathrm{uni}}$ for all $\mu \geq 4$.

Table: Optimal vs. Uniform Minimum Crowding Distance on Linear Fronts

Population ($\mu$)	Uniform $\Delta_{\mathrm{uni}}$	Optimal $\Delta^*$	Ratio $\Delta^* / \Delta_{\mathrm{uni}}$
4	$4/3$	$2$	$1.5$
6	$4/5$	$1$	$1.25$
8	$4/7$	$2/3$	$1.167$

4. Empirical Performance and Observed Distributions

Empirical studies show that the standard $(\mu+\mu)$ NSGA-II truncation often produces duplicated extreme points at the Pareto front boundaries due to ties, with quasi-random spread among interior points. The steady-state $(\mu+1)$ variant of NSGA-II, in which exactly one solution is replaced per iteration, yields nearly uniform spacing in the interior but still duplicates the two extremes. Neither variant achieves the clustered-overlap optimal configuration; their minimum crowding distance remains strictly suboptimal compared to the theoretical maximum (Ishibuchi et al., 24 Apr 2025).

In multi-objective benchmark problems, adoption of improved crowding distance (Chu et al., 2018) leads to consistently reduced Generalized Distance (GD) to the Pareto front and higher coverage as measured by the C-metric, without materially affecting distribution metrics such as SP (spacing) or $M_2$-star variance. The truthful crowding distance enables NSGA-II to avoid the exponential performance deterioration observed in classic crowding distance for many-objective problems, paralleling the provable efficiency of NSGA-III or SMS-EMOA (Zheng et al., 2024).

5. Limitations and Variant-Driven Enhancements

Classic crowding distance considers only per-objective neighbor proximity and can falsely indicate high diversity even when entire objective vectors cluster. This decoupling is particularly problematic in many-objective contexts, leading to NSGA-II's exponential runtime scaling (Zheng et al., 2024). The truthful crowding distance (tCD) corrects this by detecting genuine closeness in full objective space and ensuring a diverse coverage, with population sizes equal to the Pareto set size.

Additionally, NSGA-II's truncation, which removes all solutions sharing the minimum crowding distance in a single batch without reassessing, cannot optimally maximize the minimum crowding distance as shown in (Ishibuchi et al., 24 Apr 2025); removal of each candidate has non-local effects. A more sophisticated scheme—removing, at each step, the candidate whose absence most improves the minimum crowding distance—would, in principle, approach the optimal clustered-overlap pattern suggested by analytic arguments.

6. Connections to Diversity Metrics and Related EMO Algorithms

NSGA-II's crowding distance–based truncation does not correspond to an exact maximization of any global diversity metric such as hypervolume (SMS-EMOA), reference-line intersection (NSGA-III), or decomposition-based spread (MOEA/D) (Ishibuchi et al., 24 Apr 2025). This absence of a global criterion distinguishes NSGA-II: unlike the aforementioned methods, its diversity preservation is a result of local isolation metrics rather than global optimality. Improved and truthful crowding distance definitions provide a pathway to formalizing this connection, especially in extending the foundational principles to higher-dimensional spaces and more demanding diversity requirements (Chu et al., 2018, Zheng et al., 2024).