Cross-Domain Selection Hyper-Heuristics

• Cross-domain selection hyper-heuristics are adaptive frameworks that automatically select and configure low-level heuristics across diverse problem domains.
• They leverage standardized architectures like HyFlex to decouple domain specifics from algorithm design, enabling unified performance evaluation.
• Recent studies show that integrating dynamic selection, learning representations, and transformation principles significantly improves real-world optimization outcomes.

Searching arXiv for relevant papers on cross-domain selection hyper-heuristics and closely related hyper-heuristic paradigms. Searching for benchmark/framework papers and theoretical selection-hyper-heuristic analyses. Cross-domain selection hyper-heuristics aim to distill decades of research on problem-specific heuristic search algorithms into adaptable general-purpose search strategies (Sobotka et al., 2 Sep 2025). A selection hyper-heuristic is a randomised search methodology which chooses and executes heuristics during the optimisation process from a set of low-level heuristics, typically under a two-component architecture consisting of a heuristic-selection mechanism and a move acceptance rule; in the standard selection-perturbation setting, the controller repeatedly chooses $h\in H$, applies $x' \gets h(x)$, and accepts or rejects the move according to the acceptance criterion (Lissovoi et al., 2018). Cross-domain work is motivated by common software interfaces that package solution representation, instance data, objective evaluation, and low-level heuristics behind a shared control layer so that the algorithm designer can concentrate on adaptive general-purpose heuristic search algorithms rather than on domain-specific software components (Burke et al., 2011).

1. Conceptual scope and taxonomy

The formal background of selection hyper-heuristics is closely aligned with the Algorithm Selection Problem. In Rice’s framework, one has a problem space $\mathcal{P}$, an algorithm space $\mathcal{A}$, a feature space $\mathcal{F}$, a feature mapping $f(x)$, a performance mapping $P(A,x)$, and a selection mapping $S(f(x))$; the survey explicitly allows “algorithm” to mean “a system, a programme, a heuristic, a classifier or a configuration,” which makes the formalism directly applicable to portfolios of low-level heuristics (Kotthoff, 2012). The same survey distinguishes offline and online selection, single-algorithm choice and schedules, and per-portfolio versus per-algorithm models, all of which recur in hyper-heuristic design (Kotthoff, 2012).

Within heuristic search, dynamic heuristic choice can be written more explicitly. In planning, algorithm selection is expressed as $\pi_{\text{as}}:\mathcal{I}\to H$, adaptive algorithm configuration as $\pi_{\text{aac}}:\mathbb{N}_0\to H$, and dynamic algorithm configuration as $x' \gets h(x)$0, where the controller acts on the internal search state rather than only on the instance identity or time index (Speck et al., 2020). This places classical selection hyper-heuristics on the online end of the taxonomy: they do not merely choose one heuristic before the run, but repeatedly choose during the run.

Selection hyper-heuristics are distinct from generation hyper-heuristics and from newer code-generation paradigms. “Towards the Design of Heuristics by Means of Self-Assembly” studies the generation side of hyper-heuristics, where the goal is to construct heuristic composites rather than to choose among a fixed portfolio (Terrazas et al., 2010). ReEvo extends this direction by defining a Language Hyper-Heuristic as an HH variant where heuristics in $x' \gets h(x)$1 are generated by LLMs, with $x' \gets h(x)$2 over an open-ended space of executable programs; the paper is explicit that this is closest to an offline generation hyper-heuristic with adaptive feedback rather than to a classical selection HH (Ye et al., 2024).

2. Canonical architecture and benchmark infrastructure

The most influential software abstraction for cross-domain selection hyper-heuristics is HyFlex. HyFlex is organised around two abstract classes, ProblemDomain and HyperHeuristic: the problem-domain module owns solution representation, instance data, fitness evaluation, and low-level heuristics, while the hyper-heuristic owns the heuristic selection policy, move acceptance, memory strategy, adaptation logic, and stopping behaviour (Burke et al., 2011). The interface includes setMemorySize, copySolution, initialiseSolution(i), applyHeuristic(i,j,k), getHeuristicsOfType(type), getFunctionValue(i), and getBestSolutionValue(), and all domains are exposed as minimisation problems (Burke et al., 2011).

HyFlex also standardised a domain-independent vocabulary for low-level heuristic roles. The framework groups heuristics into mutational or perturbation heuristics, ruin-recreate heuristics, hill-climbing or local-search heuristics, and crossover heuristics (Burke et al., 2011). This categorisation is particularly useful in cross-domain settings because the high-level method can reason over operator classes without direct access to domain semantics.

The original HyFlex paper fully implemented four hard combinatorial problems—maximum satisfiability, one dimensional bin packing, permutation flow shop, and personnel scheduling—and provided a time-limited evaluation protocol resembling the Cross-domain Heuristic Search Challenge (Burke et al., 2011). Subsequent CHeSC-oriented work in the same line evaluates six domains and 30 competition instances—Maximum Satisfiability, Bin Packing, Personnel Scheduling, Flowshop, Travelling Salesman Problem, and Vehicle Routing Problem—and extends evaluation to three challenging real-world domains: PDPTW, minimum shift design, and bus driver scheduling (Sobotka et al., 2 Sep 2025). Because objective scales differ sharply across domains, HyFlex-style evaluation uses rank-based aggregation such as Borda count, while later CHeSC work commonly reports F1 scores against a referential benchmark set (Burke et al., 2011, Sobotka et al., 2 Sep 2025).

3. Selection mechanisms, acceptance, and transformed heuristic spaces

Practical selection hyper-heuristics differ primarily in how they score, remember, and revisit low-level heuristics. Q-EMCQ augments the Exponential Monte Carlo with Counter selector by introducing Q-learning memory over state-action pairs in which both states and actions correspond to four low-level operators—Cuckoo’s Lévy Flight Perturbation Operator, Flower Local Pollination Operator, Flower Global Pollination Operator, and Jaya Search Operator. The reward is $x' \gets h(x)$3 if the current action improves fitness and $x' \gets h(x)$4 otherwise, the learning rate follows $x' \gets h(x)$5, and after a non-improving step the next operator is chosen from the maximum-valued Q-table entry rather than by random reselection (Ahmed et al., 2020).

In multi-objective single-point search, the selection layer may operate over archive-management heuristics rather than over neighborhood operators. The MOSA/R reinforcement-learning hyper-heuristic uses probability matching over four low-level reseeding heuristics: choose the archive solution with minimum amount of domination, maximum amount of domination, largest hypervolume contribution, or largest crowding distance. Heuristic credits combine hypervolume increment, the proportion of newly generated solutions in the current Pareto front, normalization by the iteration interval, and a time-compensation factor, after which heuristic qualities are updated by exponential smoothing and converted to probabilities with a minimum exploration floor $x' \gets h(x)$6 (Cao et al., 2018).

Selection can also target mutation operators inside a multi-objective evolutionary algorithm. Sputnik records which mutation operators were used, ranks them by the change they induce in the average objective score between successive generations, and then applies either an elitist strategy—always choose the operator with the highest positive impact—or a caste strategy that samples among positive-impact operators proportionally to their impact. In both cases, $x' \gets h(x)$7 of selections remain purely random, preserving exploration and giving poor operators an opportunity to re-enter the elite group (Kateb et al., 2014).

A substantial shift in emphasis appears in the 2025 study of cross-domain performance principles. That work argues that cross-domain selection hyper-heuristics should not focus only on adaptive selection of low-level heuristics from a predefined set, but also on the composition of this set and its strategic transformations. Three principles are isolated: solution acceptance, LLH repetitions, and perturbation intensity. The transformed “virtual LLH” layer attaches Accept, Duration, and Intensity attributes to an original LLH, thereby modifying the effective search space seen by the selector (Sobotka et al., 2 Sep 2025). With an appropriately constructed transformation, a trivial unbiased random selector outperforms all available state-of-the-art hyper-heuristics on three challenging real-world domains and finds 11 new best-known solutions (Sobotka et al., 2 Sep 2025).

4. Feature spaces, domain-independent signals, and learning representations

The quality of a selection hyper-heuristic is tightly coupled to the representation used to link problem states or instances to low-level heuristics. In the nearest-rule model studied on CSPs, a selector consists of rules, each defined by a condition vector in feature space and an attached low-level heuristic; the heuristic associated with the closest rule is applied to the current feature vector. That line of work argues that selector quality depends strongly on feature geometry and introduces explicit transformations $x' \gets h(x)$8 and $x' \gets h(x)$9 as well as implicit kernel-induced distance metrics, especially the RBF kernel with $\mathcal{P}$0 (Amaya et al., 2018). In confirmatory knapsack experiments, the kernel-based selector reduced the standard deviation of profit from $\mathcal{P}$1 to $\mathcal{P}$2 on 50-item instances and from $\mathcal{P}$3 to $\mathcal{P}$4 on 100-item instances, with a significant Wilcoxon result on the harder set ($\mathcal{P}$5) (Amaya et al., 2018).

A related but non-selector line in planning investigates domain-independent heuristic representations. “Towards learning domain-independent planning heuristics” uses features such as number of multi-valued variables, quartiles of variable domain sizes, number of goal conjuncts, number of unsatisfied goal conjuncts, CEA heuristic value, FF heuristic value, number of delete effects ignored by FF’s relaxed plan, and number of operators in the FF relaxed plan (Gomoluch et al., 2017). The evaluation is explicitly limited: only three domains—Transport, Woodworking, and Parking—are used, all are represented in the training data, and no completely unseen target domain is tested (Gomoluch et al., 2017). The study nevertheless shows why shared feature spaces are a prerequisite for any stronger claim of cross-domain reuse.

Dynamic heuristic selection can replace static instance features with internal search dynamics. In planning, the controller state is built from $\mathcal{P}$6, $\mathcal{P}$7, $\mathcal{P}$8, $\mathcal{P}$9, and $\mathcal{A}$0 for each heuristic’s open list, together with the current expansion step $\mathcal{A}$1, and the policy uses differences between successive time steps rather than raw values (Speck et al., 2020). This suggests a representation strategy in which cross-domain portability comes from planner-internal statistics rather than from domain semantics.

5. Theoretical foundations

Rigorous runtime analysis has made the selection-hyper-heuristic literature unusually precise about the conditions under which simple controllers do or do not learn. On LeadingOnes, the standard Simple Random, Permutation, Greedy, and Random Gradient hyper-heuristics show “no signs of learning,” and uniform Simple Random has expected runtime $\mathcal{A}$2, which is worse than pure $\mathcal{A}$3 with runtime $\mathcal{A}$4 (Lissovoi et al., 2018). The Generalised Random Gradient hyper-heuristic changes the notion of success from one successful iteration to a learning period of length $\mathcal{A}$5; if $\mathcal{A}$6 and $\mathcal{A}$7, GRG with $\mathcal{A}$8 satisfies $\mathcal{A}$9, matching the best possible runtime achievable with those heuristics up to lower-order terms (Lissovoi et al., 2018). The same paper proves that with $\mathcal{F}$0 operators, GRG again matches the best possible asymptotic runtime when $\mathcal{F}$1 and $\mathcal{F}$2, whereas Simple Random, Permutation, Greedy, and Random Gradient worsen as $\mathcal{F}$3 increases (Lissovoi et al., 2018).

Adaptive Random Gradient removes the need to set $\mathcal{F}$4 manually. With $\mathcal{F}$5, $\mathcal{F}$6, $\mathcal{F}$7, and $\mathcal{F}$8, ARG increases or decreases $\mathcal{F}$9 multiplicatively depending on whether the current heuristic achieves at least $f(x)$0 improvements within the current learning period. The main theorem gives $f(x)$1, and the proof implies that the hyper-heuristic uses the optimal neighbourhood size in a $f(x)$2 fraction of iterations (Doerr et al., 28 May 2026).

Dynamic algorithm configuration broadens the theoretical frame beyond neighborhood-size control. In planning, Proposition 1 states that for each algorithm selection policy $f(x)$3 and adaptive algorithm configuration policy $f(x)$4 there is a dynamic control policy $f(x)$5 that expands at most as many states until a plan is found, and Theorems 1 and 2 construct families in which GBFS with dynamic control expands exponentially fewer states than either static algorithm selection or fixed alternation (Speck et al., 2020). On multimodal pseudo-Boolean benchmarks, however, the transferability of a move-acceptance hyper-heuristic depends on the variation operator. MAHH inherits the previously known $f(x)$6 behavior on CLIFF, but on $f(x)$7 with one-bit mutation its expected runtime is at least $f(x)$8 for $f(x)$9; replacing one-bit mutation by standard bit-wise mutation yields $P(A,x)$0, which is never asymptotically worse than simple elitist EAs and is asymptotically better for larger $P(A,x)$1 (Doerr et al., 2024).

6. Empirical performance, limitations, and neighboring paradigms

Cross-domain empirical evidence is strongest where the same high-level method is evaluated under the same benchmark protocol across multiple domains. The 2025 transformation study reports that NHH improves across CHeSC from $P(A,x)$2 to $P(A,x)$3, LUBY from $P(A,x)$4 to $P(A,x)$5, MC from $P(A,x)$6 to $P(A,x)$7, LGIHH from $P(A,x)$8 to $P(A,x)$9, FSILS from $S(f(x))$0 to $S(f(x))$1, TSILS from $S(f(x))$2 to $S(f(x))$3, and EAILS from $S(f(x))$4 to $S(f(x))$5 (Sobotka et al., 2 Sep 2025). In the original CHeSC competition, NHH$S(f(x))$6 and LUBY$S(f(x))$7 would rank 2nd and MC$S(f(x))$8 would rank 1st, while the same transformation principles on real-world domains produce 11 new best-known solutions, specifically 3 in PDPTW and 8 in minimum shift design (Sobotka et al., 2 Sep 2025).

At the same time, several adjacent literatures stop short of full cross-domain selection. “Towards learning domain-independent planning heuristics” explicitly restricts experimentation to three domains all represented in training and provides no unseen-domain evaluation (Gomoluch et al., 2017). “Learning Heuristic Selection with Dynamic Algorithm Configuration” deliberately uses domain-independent search statistics, but the authors state that the policies considered are not domain-independent and that the study is limited to domain-wise dynamic control policies (Speck et al., 2020). This suggests that reusable representations are further developed than zero-shot domain transfer.

Broader hyper-heuristic research is also expanding the taxonomy beyond classical selection. ReEvo reuses the same reflective evolutionary loop across six combinatorial optimisation problems and five heterogeneous algorithmic types, but it is an offline generation hyper-heuristic over executable code rather than an online selector over a fixed LLH portfolio (Ye et al., 2024). Population-based policy search for the integrated procurement and perishable inventory problem performs an offline selection-and-configuration search over item-level policy bundles $S(f(x))$9, showing that composite policy construction can outperform uniform global policies, yet it is application-specific rather than cross-domain in the classical sense (Felizardo et al., 2 Nov 2025). This suggests that the contemporary research frontier is no longer only about making the selector more adaptive; it is also about designing the heuristic space, the feedback signal, and the representation on which selection operates.