Dominated Novelty Search (DNS)

• DNS is a family of evolutionary algorithms that combine behavioral novelty, local competition, and dominance relations to effectively explore and exploit high-dimensional search spaces.
• The approach employs dynamic fitness transformations and Pareto-dominance criteria to unify fitness and novelty, eliminating the need for explicit archives or discretization.
• DNS enhances performance in continuous control and neural architecture search by reliably maintaining diverse, high-quality solutions and mitigating premature convergence.

Dominated Novelty Search (DNS) refers to a family of evolutionary algorithms that synthesize principles of local competition, behavioral novelty, and dominance relations to drive diverse, high-performing solution discovery in Quality-Diversity (QD) and related optimization problems. Contrasting with standard novelty search and multi-objective optimization, DNS introduces mechanisms by which individuals are scored based both on their behavioral uniqueness and on their relation to higher-quality (fitter or non-dominated) solutions, often via explicit domination-based criteria or fitness transformations. Recent work formalizes DNS in several algorithmic guises, addressing key limitations of archive/grid-based QD algorithms, threshold-driven local competition, and Pareto-based diversity maintenance (Bahlous-Boldi et al., 1 Feb 2025, Vo et al., 2024, Meyerson et al., 2017).

1. Historical and Theoretical Context

The genesis of DNS can be traced to the effort to resolve the premature convergence and limited exploration of canonical evolutionary algorithms. Classic novelty search eschews objective performance, rewarding behavioral outliers to maximize diversity, yet risks incurring computational waste by exploring unpromising regions of the solution space. Local competition in QD and archive-based approaches (MAP-Elites, Threshold-Elites) attempted to retain diverse, high-quality solutions but required externally-imposed grids, behavioral archives, or threshold parameters.

DNS emerges as an alternative that formalizes local competition through dynamic, domination-based transformations of the fitness or selection process. In some variants, such as "behavior domination" (Meyerson et al., 2017), novelty and quality are combined using a domination relation defined over both fitness and behavior space. Alternative approaches cast DNS as genetic algorithms with fitness transformations that reward novelty only in directions not already covered by fitter solutions (Bahlous-Boldi et al., 1 Feb 2025), or as Pareto-dominance guided novelty search in multi-objective NAS (Vo et al., 2024).

2. Core Algorithms and Mathematical Formulation

DNS via Dominated-Novelty Fitness Transformation

DNS can be implemented as a genetic algorithm where the key selection step is mediated by a dynamic fitness transformation. Let $N$ denote the population size, with individuals $x_i$ possessing fitness $f_i$ and descriptor $d_i \in \mathbb{R}^D$. For each $i$, define the set of strictly fitter solutions $D_i = \{ j \mid f_j > f_i \}$. The DNS competition fitness (dominated novelty score) is given by:

• If $|D_i| = 0$, $\tilde{f}_i = +\infty$ (individual is globally best).
• Else, $$\tilde{f}_i = \frac{1}{k} \sum_{j \in \mathcal{K}_i} \|d_i - d_j\|_2$$, where $\mathcal{K}_i$ indexes the $x_i$0 nearest neighbors among $x_i$1.

Selection proceeds by retaining the top $x_i$2 individuals by $x_i$3. This transformation unifies raw fitness and local behavioral novelty relative to strictly better solutions and obviates the need for explicit archives or discretization (Bahlous-Boldi et al., 1 Feb 2025).

Behavior Domination Partial Order

In the behavior domination framework (Meyerson et al., 2017), solutions are evaluated under a partial ordering defined as

$x_i$4

where $x_i$5 is fitness, $x_i$6 is an $x_i$7-dimensional behavior characterization, and $x_i$8 is a scaled $x_i$9 norm. $f_i$0. Non-dominated sorting under this order maintains a diverse front of stepping stones spanning distinct behavioral regions and fitness levels.

Pareto Dominance-based Novelty Search in NAS

In multi-objective NAS, DNS is realized through a Pareto-dominance relation over a set of $f_i$1 metrics $f_i$2. An elitist archive $f_i$3 of non-dominated individuals is maintained. For each candidate $f_i$4, with descriptor $f_i$5, the novelty score is:

$f_i$6

with a sign-adjustment: positive if $f_i$7, negative otherwise. Selection maximizes $f_i$8, ensuring that only non-dominated (Pareto-optimal) novelties are promoted (Vo et al., 2024).

3. Comparative Properties and Key Differences Across Variants

Property	DNS (fitness transformation) (Bahlous-Boldi et al., 1 Feb 2025)	Behavior Domination (Meyerson et al., 2017)	MTF-PDNS (Pareto DNS) (Vo et al., 2024)
Archive/Container	None (implicit via dynamic distances)	Archive for novelty, maintained front	Explicit elitist non-dominated archive
Descriptor tuning	No descriptor bounds, no threshold $f_i$9	$d_i \in \mathbb{R}^D$0 may need tuning; adaptive possible	Typically normalized metrics
Hyperparameters	Only $d_i \in \mathbb{R}^D$1 (neighbor count)	$d_i \in \mathbb{R}^D$2 (novelty/fitness weight), $d_i \in \mathbb{R}^D$3, $d_i \in \mathbb{R}^D$4	$d_i \in \mathbb{R}^D$5 (for $d_i \in \mathbb{R}^D$6-NN novelty, if used)
Scalability (high-D)	Good, no curse of dimensionality via grids	Robust up to moderate scales	High (metrics-based descriptors)
Selection mechanism	Top-$d_i \in \mathbb{R}^D$7 by dominated-novelty score	Non-dominated sort, crowding, novelty fill	Top by novelty on non-dominated set

DNS notably removes the need for grid discretization (cf. MAP-Elites), avoids fragile threshold tuning required by archive-based local competition, and adapts dynamically to the actually reachable descriptor space. In the behavior domination approach, the weighted behavioral metric is integrated directly with fitness, formalizing trade-offs in exploration and exploitation.

4. Empirical Evaluation and Performance Characteristics

DNS has been benchmarked against MAP-Elites, Cluster-Elites, and Threshold-Elites on continuous control (Walker, Ant, Ant+Blocks), high-dimensional behavior descriptors, and unsupervised QD tasks. In continuous control environments, DNS consistently outperforms TE and CE ($d_i \in \mathbb{R}^D$8), and often surpasses ME, with robustness in high-dimensional and unstructured spaces. Coverage and QD-score metrics demonstrate superior maintenance of diverse, high-performing niches. Ablation studies on $d_i \in \mathbb{R}^D$9 reveal stability across a range of locality values (Bahlous-Boldi et al., 1 Feb 2025). In neural architecture search, MTF-PDNS achieves lower IGD$i$0 and higher hypervolume than NSGA-II style baselines, with broader Pareto front coverage and reduced computational cost (Vo et al., 2024).

In stepping stone discovery and deceptive landscapes, the behavior domination DNS variant maintains diverse high-fitness solutions over more generations and across more difficult optimization settings than pure novelty or fitness-based strategies (Meyerson et al., 2017). For example, on the Exponential Focus domain, DNS reaches significantly higher fitness levels ($i$1 vs. $i$2 for pure novelty).

5. Algorithmic Implementation and Scalability

All DNS variants involve $i$3 per-generation complexity due to pairwise computation among individuals (for distances or non-dominated sorting). In practice, with $i$4 and moderate $i$5, DNS remains tractable, especially with GPU acceleration or approximate neighbor search. The only principal hyperparameter in (Bahlous-Boldi et al., 1 Feb 2025) is $i$6 (neighbor count), and empirical results show minimal sensitivity to this choice across practical ranges. In behavior domination, an additional weight $i$7 governs the trade-off and may be adapted dynamically.

Implementation practices for DNS include: using normalized descriptors commensurate with fitness scale, maintaining elitist archives where relevant, and, in the pure DNS (fitness transformation) variant, simply replacing the QD archive/grid with the DNS scoring step in existing frameworks (Bahlous-Boldi et al., 1 Feb 2025, Meyerson et al., 2017).

6. Preferred Use Cases and Practical Guidelines

DNS is most advantageous in scenarios where the descriptor space is unbounded, high-dimensional, learned, or discontinuous, making grid or threshold-based partitioning infeasible or inefficient. It excels in maintaining diversity without manual parameter tuning and adapts robustly to arbitrary descriptor geometries. Domains with nontrivial or nonstationary behavior embeddings (e.g. unsupervised or learned embeddings) particularly benefit from DNS, as do applications in neural architecture search where dominance and novelty trade-offs arise naturally in multi-metric settings (Vo et al., 2024, Bahlous-Boldi et al., 1 Feb 2025).

Practical guidelines recommend setting $i$8 in the range $i$9 for the fitness transformation approach, using adaptive or domain-informed $D_i = \{ j \mid f_j > f_i \}$0 in behavior domination DNS, and initializing descriptors to be numerically compatible with fitness differences. For archival strategies, non-dominated solutions and novelty maintenance via crowding or nearest neighbor measures are suggested.

7. Theoretical Guarantees and Limitations

DNS inherits theoretical guarantees from the underlying multi-objective or non-dominated sorting frameworks: front monotonicity, coverage guarantees under modest sampling assumptions, and explicit diversity preservation via crowding-distance or novelty-based selection (Meyerson et al., 2017). A notable benefit is the reliable maintenance of stepping stones and the avoidance of collapse onto local optima typical in pure fitness or naive novelty strategies.

A potential limitation is $D_i = \{ j \mid f_j > f_i \}$1 scaling, which can become a bottleneck for very large population sizes unless mitigated with approximate neighbor search or sparsification. DNS also presumes that behavioral descriptors or performance metrics are meaningful for distinguishing and rewarding novelty; accordingly, poor metric choice or misaligned descriptors may blunt its efficacy.

---

DNS establishes a robust, theoretically-founded methodology for integrating diversity- and quality-seeking pressures in evolutionary search and QD. Across formulations—dynamic fitness transformations, behavior domination, and Pareto-guided novelty—DNS consistently achieves superior diversity maintenance, scalability, and performance in both classic and modern high-dimensional tasks (Bahlous-Boldi et al., 1 Feb 2025, Vo et al., 2024, Meyerson et al., 2017).