Bayesian Simulated Annealing — Global Optimization with Bayesian Priors
Bayesian Simulated Annealing — Probabilistic global optimization with Bayesian priors on the energy landscape · Also known as: BSA, Bayesian SA, Bayesian Stochastic Annealing, Bayesian Thermodynamic Optimization
Bayesian Simulated Annealing (BSA) integrates Bayesian prior knowledge about the objective landscape into the simulated annealing search process. By encoding beliefs about promising regions as prior distributions and updating them as the search progresses, BSA focuses computational effort on high-probability areas of the solution space, accelerating convergence and improving solution quality compared to uninformed SA.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian Simulated Annealing when optimizing expensive black-box objective functions where prior knowledge about the solution space is available, and when standard SA converges too slowly or wastes evaluations. It is particularly effective for engineering design, hyperparameter tuning, and Bayesian inverse problems with complex multimodal landscapes. Avoid it when the objective function is cheap to evaluate (plain SA or grid search suffices), when no meaningful prior information exists (BSA degrades to standard SA with overhead), or when the posterior update step itself is computationally prohibitive relative to the objective evaluation cost.
Strengths & limitations
- Incorporates prior domain knowledge, directing the search toward promising regions and reducing unnecessary evaluations.
- Inherits SA's ability to escape local optima through probabilistic acceptance of worse solutions.
- Provides a principled probabilistic framework that yields uncertainty estimates alongside the optimal solution.
- Adaptable to both continuous and discrete optimization problems with appropriate proposal distributions.
- Posterior updates can be cached and reused across related optimization runs, improving efficiency in repeated tasks.
- Posterior update steps (especially with surrogate models) add computational overhead that can outweigh benefits for cheap objective functions.
- Requires specification of a prior distribution; a poorly chosen prior can bias the search away from the true optimum.
- Convergence proofs from standard SA do not automatically transfer when the acceptance criterion is modified by Bayesian weights.
- Implementation complexity is significantly higher than plain simulated annealing, increasing development and debugging effort.
Frequently asked
How does Bayesian Simulated Annealing differ from standard Simulated Annealing?
Standard SA uses only the Metropolis acceptance criterion based on energy difference and temperature. BSA additionally weights acceptance by Bayesian prior (and posterior) probability, steering the search toward regions deemed likely by domain knowledge. This reduces wasted evaluations in low-quality areas of the search space.
Is Bayesian Simulated Annealing the same as Bayesian Optimization?
No. Bayesian Optimization (BO) replaces the iterative search entirely with a surrogate model and an acquisition function that selects the next evaluation point. BSA retains the SA search engine but incorporates Bayesian weighting into the acceptance probability. They can be complementary: a BO surrogate can provide the posterior used inside BSA.
What prior distribution should I use if I have no domain knowledge?
Use a non-informative prior (e.g., uniform over the feasible space). In this case BSA degenerates to standard SA with a Metropolis-Hastings flavor. If previous runs exist, an empirical prior based on observed good solutions is a practical starting point.
Does BSA guarantee finding the global optimum?
No. Like standard SA, BSA provides probabilistic convergence to the global optimum under a sufficiently slow (logarithmic) cooling schedule, but this is rarely practical. The Bayesian component improves efficiency in practice but does not restore theoretical guarantees lost when using faster cooling schedules.
When is BSA preferable to genetic algorithms or particle swarm optimization?
BSA is preferable when the objective function is expensive, meaningful prior information is available to encode, and interpretable uncertainty quantification about the solution is desired. Population-based methods like GA or PSO are better when evaluations are cheap and exploration of diverse solutions is the main concern.
Sources
- Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by simulated annealing. Science, 220(4598), 671–680. DOI: 10.1126/science.220.4598.671 ↗
- Geman, S., & Geman, D. (1984). Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 6(6), 721–741. DOI: 10.1109/TPAMI.1984.4767596 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Simulated Annealing — Probabilistic global optimization with Bayesian priors on the energy landscape. ScholarGate. https://scholargate.app/en/simulation/bayesian-simulated-annealing
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Genetic AlgorithmSimulation↔ compare
- Bayesian OptimizationOptimization↔ compare
- Genetic AlgorithmOptimization↔ compare
- Markov Chain Monte CarloSimulation↔ compare
- Simulated AnnealingOptimization↔ compare