Balanced NSGA-II Enhancements

• Balanced NSGA-II is a suite of algorithm modifications designed to address imbalance issues by refining diversity, convergence, and Pareto front distribution.
• Key variants like evoNSGA-II and OTNSGA-II employ techniques such as evolvability-based selection and orthogonal initialization to maintain balanced complexity across solutions.
• Empirical studies show these enhancements achieve lower maximal empty intervals and higher hypervolumes, outperforming classic NSGA-II on benchmark problems.

Balanced NSGA-II refers to a family of algorithmic strategies and specific modifications of the original Non-dominated Sorting Genetic Algorithm II (NSGA-II) designed to address distributional or convergence imbalances endemic to classic NSGA-II in multi-objective optimization scenarios. These imbalances include dominance of low-complexity or poorly diversified solutions, suboptimal spacing on the Pareto front, and inefficient exploitation of problem structure. Variants under this banner—including evolvability-based capping (evoNSGA-II), improved crowding-distance formulations, on-the-fly crowding distance updates, steady-state survivor replacement, orthogonal-initialization with adaptive cluster-pruning (OTNSGA-II), and others—apply population control or selection scheme refinements that provably or empirically lead to more evenly distributed, better-converged fronts.

1. Motivations for Balancing in NSGA-II

The original NSGA-II, as established by Deb et al., combines fast non-dominated sorting with a crowding-distance-based diversity preservation, and generational survivor selection. While effective, NSGA-II exhibits the following deficiencies across a variety of domains:

• Takeover by Low-Complexity, Low-Evolvability Solutions: In symbolic regression via GP, smaller trees often proliferate disproportionately but are rarely fertile in producing improved offspring, causing rapid evolvability degeneration and slow progress along the accuracy–complexity Pareto front (Liu et al., 2022).
• Poor Pareto Coverage With Small Populations: When the Pareto front cardinality substantially exceeds the population, selection and elimination steps may leave large empty intervals or “gaps” on the true front (Zheng et al., 2022).
• Crowding-Distance Limitations: The symmetric, hypercube-based crowding distance in classic NSGA-II yields “flat plateaus” where all solutions inside a box have equal survival priority, irrespective of proximity to the true front (Chu et al., 2018).
• Initialization and Overcrowding Effects: Random initial populations and simplistic survivor selection can cause clustering, premature convergence, and uneven solution spread, impeding both diversity and convergence (Yang et al., 2019).

Balanced NSGA-II variants target these issues using algorithmic controls tailored to maintain population diversity, Pareto locality, and effective use of the front's structure.

2. Evolvability-Based NSGA-II (evoNSGA-II)

EvoNSGA-II introduces a meta-selection layer that explicitly tracks and exploits the “evolvability” of individuals by complexity (size). Here, evolvability $e(s)$ for solutions of size $s$ is the empirical probability that an offspring generated by variation (crossover or mutation) from a parent of size $s$ improves on a baseline accuracy (typically the population median). The algorithm maintains per-size statistics:

• $\mathrm{attempts}[s]$: offspring generated from parents of size $s$
• $\mathrm{successes}[s]$: such offspring exceeding median accuracy
• $$e(s) = \frac{\mathrm{successes}[s]}{\mathrm{attempts}[s]}$$

For survivor selection, NSGA-II's non-domination and crowding hierarchy is augmented: a per-size cap $B[s]$ is computed by normalizing $e(s)$ such that $\sum_{s=1}^{S_{\max}} B[s] = N$. Only up to $B[s]$ individuals of size $s$ are allowed to survive in the population. Sizes unseen in a generation are assigned interpolated evolvabilities. This policy prevents low-evolvability, typically small solutions from over-replicating, maintaining a balanced complexity distribution.

Empirical evaluations on ten real-world regression datasets confirm that evoNSGA-II consistently outperforms classical NSGA-II, SPEA2, and several other variants in hypervolume coverage, especially for larger population sizes. Only evoNSGA-II and NSGA-II+PD reliably discovered large, high-quality solutions; evoNSGA-II generally achieved higher population hypervolumes (Liu et al., 2022).

3. Crowding-Distance Improvements and Gradient Bias

Chu & Yu propose an improved crowding-distance for NSGA-II, intending to induce a convergence bias toward the Pareto front while retaining spread. In the standard NSGA-II, for each individual $j$ in a front (of size $n$), the crowding distance is

$$d_j^{\text{orig}} = \sum_{k=1}^{m} \frac{f_{j+1}^k - f_{j-1}^k}{f^{k}_{\max} - f^k_{\min}}$$

where $f^k$ indexes objectives, and individuals are sorted by $f^k$.

The modified version instead uses a forward difference: $$d_{j}^{\text{new}} = \sum_{k=1}^m \frac{f_{j+1}^k - f_j^k}{f^k_{\max} - f^k_{\min}}$$ Boundary points retain infinite distance. This generates a “sloping” surface in diversity space, systematically favoring individuals closer to the front tip.

Experimental results on nine benchmarks (ZDT, SCH, FON, etc.) demonstrate that improved crowding distance yields lower generalized distance to the Pareto front, higher binary coverage, and equal or superior spread. The computational cost remains $O(mn\log n)$ per front, introducing no new hyperparameters (Chu et al., 2018).

4. Dynamic Survivor Control: On-the-Fly and Steady-State NSGA-II

To counteract static crowding-distance artifacts and suboptimal front coverage, two dynamic survivor strategies have been theoretically and empirically validated:

• On-the-fly (Incremental) Crowding Distance Update: During truncation of the last filled front, crowding distances are re-computed after each removal. A priority queue and doubly-linked lists track immediate neighbors in each objective, updating local distance as individuals are excised. This ensures removal decisions always reflect the current local density (Zheng et al., 2022).
• Steady-State NSGA-II: Instead of generational replacement, each iteration produces a single offspring; survivor elimination is executed immediately by removing the worst (by dominance rank and crowding distance) from the temporary $(N+1)$-sized population.

Both approaches lead to provably tighter control of the maximal empty interval ($\text{MEI}$) on the Pareto front. Specifically, the $\text{MEI}$ is at worst twice the information-theoretic optimum $\lceil n/(N-1)\rceil$ (for a front of cardinality $n$), and the gap remains stable post-convergence. Experiments on the OneMinMax benchmark show classical NSGA-II displays gaps $2$--$3\times$ larger than the balanced variants (Zheng et al., 2022).

Maximal Empty Interval Table

Population Size $N$	Classic NSGA-II $\text{MEI}$	On-the-Fly/Steady-State $\text{MEI}$	Theoretical Optimum
301	(7, 8, 9)	(3, 3, 3)	3
151	(14,15,17)	(5, 5, 6)	5
76	(23,26,29)	(11,12,12)	9

Values are quartiles over 100 generations post-front convergence.

5. Orthogonal Initialization and Adaptive Cluster-Pruning (OTNSGA-II)

OTNSGA-II improves NSGA-II's population distribution by optimizing two stages:

• Orthogonal-Experiment Initialization: Instead of random sampling, initial individuals are generated via an orthogonal experimental design in the dominant variable dimension, ensuring a wide, uniform coverage of decision/objective space. Self-Adaptive Orthogonal Crossover (SOC) produces offspring by systematically recombining parent solution segments via orthogonal arrays.
• Adaptive Clustering-Pruning: At each generation, the $2N$ merged population is clustered (e.g., via $k$-means) in an extended feature space that includes decision variables, objectives, dominance rank, and crowding distance. Within each cluster, redundant individuals—assessed by intra-class similarity—are pruned according to an adaptive fraction $r_k = 1 - \delta P_k$, where $P_k$ is the average similarity and $\delta$ is a tunable parameter. Outliers far from the true front are also removed.

Performance metrics (GD, IGD, hypervolume, etc.) demonstrate OTNSGA-II yields average improvements of $\sim$90% in convergence and $\sim$40% in distribution metrics across ZDT, DTLZ, and UF benchmarks relative to classic NSGA-II (Yang et al., 2019).

6. Comparative Algorithmic Features

Variant	Main Balancing Mechanism	Empirical/Theoretical Impact
evoNSGA-II	Evolvability-based complexity cap	Balanced size diversity, higher HV (Liu et al., 2022)
Improved CD	Forward-difference crowding distance	Faster convergence, improved spread (Chu et al., 2018)
On-the-fly CD NSGA-II	Incremental survivor elimination	Minimizes Pareto gap, provable bounds (Zheng et al., 2022)
Steady-State NSGA-II	One-by-one survivor update	Minimizes Pareto gap, low MEI (Zheng et al., 2022)
OTNSGA-II	Orthogonal design + clustering-pruning	Superior initial/external diversity (Yang et al., 2019)

7. Limitations and Further Directions

Balanced NSGA-II approaches incur moderate additional overhead versus classic implementations, especially for cluster-based or incremental survivor selection mechanisms. Estimates for evolvability can be noisy in small populations, and some methods require hyperparameter selection (e.g., pruning threshold $\delta$ or number of clusters $k$). Potential improvements include dynamic parameter setting, integration of hypervolume-based selection in initialization and pruning, and parallelization of computationally intensive stages (Yang et al., 2019).

Direct theoretical comparison between approaches is mostly available for Pareto front coverage scenarios; performance on high-dimensional, real-world problems remains an active subject for empirical study.

Balanced NSGA-II methods represent a systematic portfolio of enhancements that, by controlling selection dynamics at the levels of diversity, convergence, and population structure, achieve demonstrably stronger performance for multi-objective optimization in both benchmark and practical contexts.