Sörensen, Sevaux, and Glover set out to write not a list of methods but the story of how a field changed its mind about itself (Sörensen, Sevaux, & Glover, 2018). Having one of the authors be Fred Glover, who coined the word metaheuristic in the first place, gives the account unusual authority. We read it as a working practitioner who lived through its middle chapters — presenting tabu search at the EURO conference in Glasgow in 1994 and corresponding with Glover as the framework took hold. Three of its arguments deserve wider circulation, and the first is a matter of definition.
A framework, not an algorithm
The history adopts a precise definition: a metaheuristic is a high-level, problem-independent framework offering strategies from which heuristic algorithms are built, rather than a fixed sequence of steps (Sörensen & Glover, 2013). The distinction is not pedantic. Variable neighbourhood search, on this reading, is the idea that switching local search operators when one stalls will reach solutions a single operator cannot, since a local optimum for one neighbourhood is rarely a local optimum for another (Mladenović & Hansen, 1997). Naming that idea an algorithm misses the point. A practitioner still has to instantiate it for the problem at hand, and the engineering involved in that step is where solution quality is won or lost. Recognising metaheuristics as frameworks rather than recipes is the conceptual shift the authors place at the centre of the field’s maturation — the move from a method-centric era to a framework-centric one.
Memory over randomness
Glover’s deeper argument concerns what gives heuristic search its power. The field developed an early and lasting attraction to randomness, the intuition that useful order could emerge from blind perturbation given enough restarts. Tabu search advanced a different premise: that information gathered during a search should be remembered and used to guide it, through intensification toward regions that have yielded good solutions and diversification into regions left unexplored (Glover, 1989). The history is careful to separate this from the exploitation-and-exploration language borrowed from control theory, where a memoryless recipe runs until it fails and random departures follow. Intensification and diversification are memory-based and purposeful, moving deliberately into new territory rather than shaking the system and hoping (Glover & Laguna, 1997). Forty years of results have favoured the memory view. A metaheuristic that exploits the structure of its problem reliably beats one that treats the problem as a black box, whatever framework either is built on.
The misuse of the name
A definition matters most when it is abused, and the history devotes a pointed section to a parallel development it declines even to number among the field’s real periods. Beginning in the 1980s and continuing still, a stream of methods has appeared, each motivated by a fresh metaphor — ants and bees giving way to wolves, fireflies, mine blasts, and cloud formation — and each claiming the status of a novel metaheuristic. Many of these metaphors describe no optimising process at all, and the mechanics underneath are usually familiar ones in unfamiliar costume. The authors are blunt that this is not science, an assessment argued at length elsewhere (Sörensen, 2015). The definitional stakes are real. If any metaphor can license a new metaheuristic regardless of whether it optimises anything, the term describes nothing. Holding the framework definition is what keeps the word meaningful.
Toward a science
The history closes by naming a period that has not yet arrived: a scientific period in which the design of metaheuristics becomes a discipline rather than a craft. Glover noted as early as 1977 that exact algorithms enjoyed academic prestige while heuristics were treated as expedient and accorded lower standing, and the imbalance in theoretical underpinning has been slow to correct. The authors set out what the transition requires: honest testing protocols, meta-analysis to establish whether celebrated components earn their complexity, disclosure of source code, and general-purpose heuristic solvers built on accepted modelling languages. Glover’s own recent direction points the same way, with multi-wave algorithms that exploit how the reliability of search information grows as decisions accumulate, and a renewed emphasis on joining adaptive memory with classical learning (Glover, 2016). Our reading, after three decades alongside the field, is that the framework definition and the primacy of memory are settled gains, and the work that remains is to make the design of these methods answer to evidence. A field that can write its own intellectual history this clearly is closer to that science than the history modestly claims.