A Review on Single-Problem Multi-Attempt Heuristic Optimization

In certain real-world optimization scenarios, practitioners are not interested in solving multiple problems but rather in finding the best solution to a single, specific problem. When the computational budget is large relative to the cost of evaluating a candidate solution, multiple heuristic alternatives can be tried to solve the same given problem, each possibly with a different algorithm, parameter configuration, initialization, or stopping criterion. In this practically relevant setting, the sequential selection of which alternative to try next is crucial for efficiently identifying the best possible solution across multiple attempts. However, suitable sequential alternative selection strategies have traditionally been studied separately across different research topics and have not been the exclusive focus of any existing review. As a result, the state-of-the-art remains fragmented for practitioners interested in this setting, with surveys either covering only subsets of relevant strategies or including approaches that rely on assumptions that are not feasible for the single-problem case. This work addresses the identified gap by providing a focused review of single-problem multi-attempt heuristic optimization. It brings together suitable strategies for this setting that have been studied separately through algorithm selection, parameter tuning, multi-start, and resource allocation. These strategies are described using a unified terminology within a common framework, which supports the construction of a taxonomy for systematically organizing and classifying them. The resulting comprehensive review facilitates both the identification and the development of strategies for the single-problem multi-attempt setting in practice.

💡 Research Summary

This paper addresses a practical yet under‑studied scenario that the authors name “Single‑Problem Multi‑Attempt Heuristic Optimization” (SIMHO). In many real‑world applications the evaluation of a candidate solution is cheap (seconds) while the total computational budget is large, allowing a practitioner to run several heuristic attempts on the same optimization instance. Each attempt may differ in four dimensions: the underlying heuristic algorithm, its parameter configuration, the initialization of the search, and the stopping criterion. The central question is how to select, in a sequential manner, which alternative to try next so that the best possible solution is found within the available budget.

The authors observe that the literature already contains many strategies that can be interpreted as “alternative‑selection” mechanisms, but these strategies are scattered across four traditionally separate research topics: algorithm selection, parameter tuning, multi‑start (restart) strategies, and resource‑allocation (budget‑allocation) methods. Moreover, existing surveys either focus on only a subset of these strategies or include approaches that rely on assumptions (e.g., a large set of previously solved instances) that are not feasible when only a single problem instance is at hand. Consequently, practitioners interested in SIMHO lack a unified view of the state‑of‑the‑art.

To fill this gap, the paper makes three major contributions. First, it formally defines the SIMHO problem as a sequential decision‑making task. The formalism consists of a finite set of alternatives (A = {a_1,\dots,a_K}) (each a specific combination of algorithm, parameters, initialization, and stopping rule), a total evaluation budget (B), a policy (\pi) that decides which alternative to evaluate next based on the history of observed performances, and a feedback function (f(a,t)) that returns the quality of the solution obtained after the (t)-th attempt of alternative (a). Second, the authors introduce a unified terminology and a common abstraction that maps all existing SIMHO‑relevant methods onto this formalism. Under this abstraction, classic restart strategies (e.g., uniform random restarts of Nelder‑Mead or CMA‑ES), portfolio‑based budget distribution (e.g., the work of Souravlias et al.), successive‑halving or racing procedures, and Bayesian‑optimization‑based selection are all special cases of the same sequential allocation problem. Third, they construct a comprehensive taxonomy that classifies strategies along four orthogonal dimensions: (i) selection criterion (predicted performance, uncertainty, cost‑effectiveness), (ii) exploration‑exploitation mechanism (global exploration, local restart, adaptive allocation), (iii) information source (online learning from the current run, offline models from other instances, hybrid), and (iv) execution environment (single‑core, parallel, cloud‑based). This taxonomy makes explicit which components a given method controls and which it leaves implicit, thereby enabling systematic comparison and hybridization.

The paper also surveys related work, distinguishing between single‑problem strategies (which rely solely on information gathered during the current optimization) and multi‑problem strategies (which exploit data from a family of previously solved instances). It argues that multi‑problem methods such as per‑instance algorithm selection, offline hyperparameter tuning, and meta‑learning are generally unsuitable for SIMHO because they require a costly pre‑training phase and a representative instance set that may not exist. Hyper‑heuristics, which dynamically compose low‑level operators during a single run, are classified as single‑problem strategies and thus compatible with SIMHO; the authors discuss credit‑assignment, selection, and move‑acceptance components of hyper‑heuristics in this context.

Although the paper does not present extensive empirical evaluations, it outlines how the unified framework can be instantiated in practice. For example, an initial phase may sample a diverse set of alternatives uniformly; subsequent phases use Bayesian updating (e.g., Thompson Sampling) to bias the sampling toward promising alternatives, while a successive‑halving schedule eliminates poorly performing configurations and reallocates budget to the survivors. Such hybrid policies combine the strengths of racing (fast elimination) and Bayesian optimization (probabilistic modeling of performance) and are argued to be especially effective when the budget is limited but the evaluation cost is low.

The discussion highlights several avenues for future research. Extending the framework to parallel or cloud environments raises questions about synchronization overhead and dynamic budget scaling. Incorporating multi‑objective considerations, handling noisy or stochastic objectives, and designing adaptive stopping criteria that react to observed convergence rates are identified as open challenges. The authors also note that a thorough experimental benchmark—covering continuous, combinatorial, and simulation‑based problems—would be essential to validate the taxonomy and to quantify the benefits of different policy choices.

In conclusion, the paper provides the first dedicated review that unifies algorithm‑selection, parameter‑tuning, restart, and resource‑allocation strategies under a single formalism tailored to the SIMHO scenario. By offering a clear taxonomy and a common language, it equips both researchers and practitioners with a roadmap for selecting, comparing, and designing sequential alternative‑selection policies that can efficiently exploit abundant computational resources when solving a single, high‑stakes optimization problem.