Skip to content

Optimization by Simulated Annealing

Why this mattered

Kirkpatrick, Gelatt, and Vecchi’s paper mattered because it reframed hard optimization problems as physical systems exploring an energy landscape. Instead of treating local minima as failures to be avoided by problem-specific tricks, simulated annealing made them part of a general search process: allow uphill moves at a temperature-controlled probability, then gradually reduce that temperature so exploration gives way to exploitation. The paradigm shift was not a new exact algorithm for NP-hard problems, but a portable way to reason about optimization when exhaustive search was impossible and greedy descent was too brittle.

What became newly possible was a broadly applicable heuristic for very large combinatorial design spaces, especially those with many competing constraints and rugged objective functions. Problems in circuit placement, routing, scheduling, graph partitioning, and physical design could now be attacked with a common recipe grounded in the Metropolis algorithm and the thermodynamic analogy of annealing. The paper also helped legitimize randomized optimization: controlled randomness was not noise around a deterministic method, but the central mechanism for escaping poor local optima and sampling alternatives.

Its influence carried into later breakthroughs in stochastic search, metaheuristics, and machine learning. Genetic algorithms, tabu search, Markov chain Monte Carlo methods, Boltzmann machines, and later probabilistic inference and energy-based models all share, in different ways, the idea that difficult computational problems can be approached by navigating or sampling high-dimensional landscapes. Simulated annealing became one of the clearest early demonstrations that concepts from statistical physics could supply not just metaphors, but working algorithms for computation.

Abstract

There is a deep and useful connection between statistical mechanics (the behavior of systems with many degrees of freedom in thermal equilibrium at a finite temperature) and multivariate or combinatorial optimization (finding the minimum of a given function depending on many parameters). A detailed analogy with annealing in solids provides a framework for optimization of the properties of very large and complex systems. This connection to statistical mechanics exposes new information and provides an unfamiliar perspective on traditional optimization problems and methods.

Sources