Neural networks are now treated as the engine of modern artificial intelligence, which makes it easy to assume they arrived fully formed around 2012. Within optimisation research, the picture is older and more interesting. Decades before deep learning reached the mainstream, the metaheuristics community already regarded neural networks as a legitimate member of the methodological family: studied, debated, hybridised, and assessed on the same empirical footing as simulated annealing, genetic algorithms, and tabu search. We think the way that community held the method, with patience rather than hype, is worth revisiting.
A serious method, honestly assessed
Optimisation gave neural networks an early and concrete role. Hopfield and Tank showed that certain combinatorial problems could be framed as energy minimisation in a recurrent network, with the network’s stable states corresponding to candidate solutions (Hopfield & Tank, 1985). Standard reference works of the period covered the approach alongside its rivals, and the leading AI textbook of the decade treated neural networks as one method among several under active development, neither dismissed nor oversold (Russell & Norvig, 1995). Within the metaheuristics community, the honest assessment on the evidence then available was that artificial neural networks had not yet proved competitive with other metaheuristics on hard combinatorial problems. That conclusion was cautious in proportion to the evidence, and it was a judgement about the present state of the method, not a verdict on its potential.
Engagement, not dismissal
Caution did not mean neglect. Practitioners engaged with neural networks as a serious research direction and looked actively for ways to combine them with established search methods. Glover’s work on optimisation by ghost image processes integrated Kohonen’s self-organising maps and the Hopfield and Tank penalty formulation with the adaptive memory structures of tabu search, treating the methodologies as complementary rather than competing (Glover, 1994). Dialogue between the two communities ran in both directions. Researchers who built their careers on evolutionary or trajectory-based methods read the neural network literature, cited it, and folded its ideas into hybrid frameworks. A method kept alive in that way is a method a field expects to need later.
What changed, and what did not
When the reversal came, it was a change in conditions rather than a change in kind. Around 2012, neural networks moved from promising to dominant across vision and language, driven by graphics-processor compute and dataset scale that earlier decades could not supply (Krizhevsky, Sutskever, & Hinton, 2012; LeCun, Bengio, & Hinton, 2015). Those architectures owed a clear debt to ideas worked out in the 1980s and 1990s. What had been missing was never the concept. Missing instead were the hardware to train large models and the data to train them on, and both arrived together. Consensus advice of the late 1990s, which read the evidence of its moment correctly, turned out to be reading a temporary state as a permanent one.
Where they sit now
Neural networks have since returned to optimisation as an active partner, much as the hybrid instinct of the 1990s anticipated. Learned models now guide combinatorial solvers, predicting promising branching decisions, representing problem instances, and steering search toward regions worth the effort (Bengio, Lodi, & Prouvost, 2021). Graph neural networks have proved a natural fit for routing, scheduling, and assignment problems whose structure is itself a graph. We read this as the original judgement vindicated on its own terms. A method can be both not-yet-competitive and worth sustained investment, and the metaheuristics community treated neural networks as exactly that for the better part of two decades.
That wider lesson outlasts the example. A field serves itself well when it assesses a promising method honestly, reports what the evidence shows, and keeps engaging with the method while the surrounding conditions catch up. We try to hold new methods the same way now, neither dismissing them on early results nor adopting them on enthusiasm alone.