Diverse-BBO: Diversity-Aware Optimization

• Diverse-BBO is a framework that replaces traditional pointwise optimization with set-valued objectives to ensure both performance and explicit diversity in the solutions.
• It employs various mathematical models such as determinantal point processes, trust-region search, and ε-optimal basin coverage to balance exploration, exploitation, and diversity.
• Algorithmic variants including diverse batching, portfolio search, and explicit acquisition design make the approach applicable in Bayesian optimization, games, reinforcement learning, and AutoML.

Diverse-BBO denotes a family of diversity-aware black-box optimization formulations in which the goal is not merely to identify a single optimum, but to construct a set of solutions that jointly satisfy high quality and explicit diversity requirements. In the supplied literature, the term appears in several closely related senses: as diverse batch Bayesian optimization via determinantal point processes, as portfolio-valued Bayesian optimization with trust regions, as diversity-aware best-response optimization in open-ended games, as quality-diversity search over reinforcement-learning agents, and as a synthetic benchmark suite for Meta-Black-Box Optimization generated by latent-space reverse engineering (Nava et al., 2021, Maus et al., 2022, Nieves et al., 2021, Pierrot et al., 2023, Wang et al., 19 Sep 2025).

1. Conceptual scope and defining objectives

A common structure across these usages is the replacement of the classical objective

$\arg\max_{x \in \mathcal{X}} f(x)$

by a set-valued objective. In ROBOT, the target is a ranked set $S^* = \{x_1^*,\dots,x_M^*\}$, where $x_1^*$ maximizes $f$, and each subsequent $x_i^*$ is the best solution that is sufficiently diverse from all higher-ranked elements under a user-specified diversity function $\delta$ and threshold $\tau$ (Maus et al., 2022). In EDU, the target is a basket of $\epsilon$-optimal solutions drawn from disjoint subregions of an $\epsilon$-optimal set $\mathcal{R}_\epsilon$, so diversity is tied to covering distinct near-optimal basins rather than maximizing a single distance statistic (Miller et al., 2024). In divTuRBO1, diversity is enforced through a distance threshold $S^* = \{x_1^*,\dots,x_M^*\}$0, and the “best diverse solution” is the best feasible point outside a forbidden neighborhood of a reference set, or, if no feasible point exists, the farthest candidate from that set (Perera et al., 2 Nov 2025).

A second recurring pattern is that diversity is treated as a set property rather than a pointwise scalar objective. In the DPP-based game-theoretic formulation, behavioural diversity is the expected cardinality of a game-specific determinantal point process,

$S^* = \{x_1^*,\dots,x_M^*\}$1

with $S^* = \{x_1^*,\dots,x_M^*\}$2, where $S^* = \{x_1^*,\dots,x_M^*\}$3 is the payoff table over strategies (Nieves et al., 2021). In DPP-BBO for batched BO, diversity enters through a subset probability proportional to a determinant,

$S^* = \{x_1^*,\dots,x_M^*\}$4

so repulsion is induced at the level of the entire batch rather than individual points (Nava et al., 2021). This suggests that Diverse-BBO is best understood as a family of optimization problems whose output is a portfolio, archive, population, or basket, and whose constraints or objectives are explicitly defined on that set.

2. Mathematical models of diversity

The literature instantiates diversity through several non-equivalent mathematical objects. ROBOT uses a user-defined symmetric diversity function $S^* = \{x_1^*,\dots,x_M^*\}$5 with hard constraints $S^* = \{x_1^*,\dots,x_M^*\}$6, yielding a ranked constrained optimization problem (Maus et al., 2022). This model is agnostic to the geometry of $S^* = \{x_1^*,\dots,x_M^*\}$7 and accommodates domain-specific similarities such as fingerprint similarity, trajectory distance, Euclidean parameter distance, or portfolio disjointness (Maus et al., 2022).

Open-ended game formulations define diversity in payoff space. The gamescape of a population $S^* = \{x_1^*,\dots,x_M^*\}$8 is the convex hull

$S^* = \{x_1^*,\dots,x_M^*\}$9

and maximizing DPP-based diversity enlarges that gamescape, giving a geometric interpretation of behavioural novelty as the addition of new payoff-space directions (Nieves et al., 2021). The associated diverse best response augments payoff with a diversity term,

$x_1^*$0

which yields uniqueness through strict concavity and convergence guarantees in the stated two-player settings (Nieves et al., 2021).

EDU instead formalizes diversity through $x_1^*$1-optimal regions. The $x_1^*$2-optimal set

$x_1^*$3

is partitioned into disjoint subregions $x_1^*$4, and the objective is to recover a representative near-optimal minimizer from each region (Miller et al., 2024). This shifts the notion of diversity from pairwise separation to basin coverage.

Quality-diversity formulations in RL and evolutionary search add yet another model. In MAP-Elites, diversity is defined by a behaviour descriptor $x_1^*$5, and the archive stores one elite per discretized cell of behaviour space. The abstract of “Evolving Populations of Diverse RL Agents with MAP-Elites” states that the framework evolves populations of agents “whose definition include hyperparameters and all learnable parameters” rather than only policies, thereby broadening the search object itself (Pierrot et al., 2023).

3. Algorithmic families in Bayesian optimization

In Bayesian optimization, Diverse-BBO has developed along three main lines. The first is diverse batching. DPP-BBO wraps any randomized batch acquisition distribution $x_1^*$6 with a DPP likelihood,

$x_1^*$7

and DPP-TS instantiates this with Thompson sampling and the mutual-information kernel $x_1^*$8 (Nava et al., 2021). The resulting Bayesian simple regret bound for DPP-TS is tighter than the corresponding batched TS bound, and the method empirically improves simple regret on synthetic tasks and Cox-process sensing (Nava et al., 2021).

The second is ranked portfolio search with trust regions. ROBOT maintains $x_1^*$9 rank-ordered trust regions $f$0, uses one global surrogate model, and reconstructs a feasible ranked set from all accumulated data by greedily enforcing $f$1 against higher-ranked incumbents (Maus et al., 2022). Its distinguishing feature is that diversity is handled as a hard feasibility condition, not as a regularizer. The method was reported to find large sets of high-performing diverse solutions while requiring few additional function evaluations compared to finding a single best solution (Maus et al., 2022).

The third is explicit acquisition design for near-optimal basin coverage. EDU defines a Diverse Improvement utility and derives a closed-form DEI acquisition under a GP surrogate. The closed form exposes an exploration-exploitation-diversity trade-off controlled by $f$2, and the batch version multiplies the summed DEI values by a factor penalizing large posterior correlations among batch elements (Miller et al., 2024). In a separate trust-region line, divTuRBO1 modifies TuRBO1 by selecting trust-region centers and new candidate points through the best-feasible-or-farthest rule relative to a reference set $f$3; sequential and interleaving compositions then build a full diverse set of size $f$4 (Perera et al., 2 Nov 2025).

4. Beyond classical BO: games, RL, and AutoML

In open-ended games, Diverse-BBO is realized as diversity-aware best-response optimization. Diverse fictitious play, diverse PSRO, and diverse $f$5-PSRO incorporate the DPP diversity metric directly into best-response objectives, prove uniqueness of the diverse best response, and establish convergence in the stated settings. A central theorem shows that maximizing the DPP-based diversity term enlarges the gamescape, which is the convex polytope spanned by payoff vectors (Nieves et al., 2021). Empirically, these methods achieve at least the same, and in most games, lower exploitability than PSRO solvers while finding effective and diverse strategies (Nieves et al., 2021).

In RL-oriented quality-diversity search, MAP-Elites is coupled with policy optimization. The abstract of (Pierrot et al., 2023) identifies the key motivation: early MAP-Elites relied exclusively on random search and was sample-inefficient for high-dimensional neural policies, while RL-coupled variants inherited hyperparameter sensitivity, stochasticity, and training instability. The proposed response is a flexible framework that can use any RL algorithm and evolves populations of agents rather than only policies (Pierrot et al., 2023). This suggests an outer quality-diversity layer over agents and an inner RL optimization process over policy parameters.

In CASH and ensemble learning, DivBO injects explicit diversity search into BO-based AutoML. It learns a diversity surrogate for pairwise diversity between unseen configurations, constructs a temporary pool through ensemble selection, and uses a weighted acquisition that combines performance rank and diversity rank (Shen et al., 2023). On 15 public datasets, DivBO achieved the best average ranks, $f$6 on validation and $f$7 on test error, among 10 compared methods (Shen et al., 2023).

5. Diverse-BBO as a benchmark suite for MetaBBO

A distinct and literal use of the name appears in the benchmark paper on latent-space reverse engineering. There, Diverse-BBO is a set of 256 synthetic problem instances generated for Meta-Black-Box Optimization (Wang et al., 19 Sep 2025). LSRE first computes 21-dimensional ELA features for a large target distribution derived from 24 CoCo-BBOB functions, five dimensions, and 270 random rotation-and-shift instances per template, then trains a fully connected autoencoder that maps those features to a 2-dimensional latent space (Wang et al., 19 Sep 2025). A $f$8 uniform grid over that latent space yields 256 target latent points, and genetic programming searches for symbolic functions whose encoded ELA features minimize latent-space $f$9 distance to those targets (Wang et al., 19 Sep 2025).

The resulting benchmark has two intended properties. First, it offers controllable, uniform, and fine-grained instance diversity. Second, it is specifically designed as a training set for MetaBBO rather than as a traditional evaluation-only benchmark (Wang et al., 19 Sep 2025). Across DEDDQN, LDE, SYMBOL, and GLEET, training on Diverse-BBO gave the best average synthetic-generalization rank, and the latent coverage analysis showed that its instances covered realistic regions better than CoCo-BBOB, MA-BBOB, or GP-BBOB (Wang et al., 19 Sep 2025).

6. Empirical behavior, limitations, and open questions

Across these lines of work, several empirical regularities recur. Diversity-aware optimization often improves robustness or coverage with limited extra cost: ROBOT can discover large sets of high-performing diverse solutions with few additional evaluations relative to single-solution BO (Maus et al., 2022); DPP-BBO improves batch simple regret (Nava et al., 2021); EDU improves coverage of $x_i^*$0-optimal basins in rover and engine-control problems (Miller et al., 2024); and divTuRBO1-seq and divTuRBO1-int perform well, particularly in larger dimensions, even with a limited evaluation budget (Perera et al., 2 Nov 2025). In games, DPP-based diversity enlarges the gamescape and is associated with lower exploitability (Nieves et al., 2021). In AutoML, explicit diversity search improves ensemble construction (Shen et al., 2023).

The main technical caveat is that Diverse-BBO is not tied to a single diversity primitive. Results depend strongly on the choice of $x_i^*$1, $x_i^*$2, behaviour descriptors, DPP kernels, or $x_i^*$3. ROBOT explicitly warns that $x_i^*$4 must be chosen so that the feasible region is non-empty and not vanishingly small (Maus et al., 2022). EDU relies on GP surrogates and a meaningful tolerance $x_i^*$5 (Miller et al., 2024). The MetaBBO benchmark paper notes that there is no standard explicit diversity metric for BBO benchmarks and introduces ASRD as an indirect measure of instance diversity (Wang et al., 19 Sep 2025). A plausible implication is that the field remains methodologically plural: the unifying theme is optimization of sets under explicit diversity structure, but the correct formalization is application-dependent.

In that sense, Diverse-BBO names both a research agenda and a toolbox. It encompasses DPP-based subset selection, trust-region portfolio search, $x_i^*$6-optimal basin discovery, archive-based quality-diversity search, diversity-aware game solving, and latent-space benchmark generation. What these variants share is a refusal to collapse black-box optimization to one incumbent solution when the downstream problem requires alternatives, robustness, or behavioural coverage.