Bayesian Genetic Algorithm — Probabilistic model-guided evolutionary optimization
Also known as: BGA, Bayesian-guided GA, Probabilistic GA, EDA-GA
A Bayesian Genetic Algorithm (BGA) replaces traditional crossover and mutation operators with a probabilistic Bayesian network learned from selected high-fitness individuals. At each generation the algorithm builds a graphical model of promising solution structure, then samples new offspring from that model, enabling the search to capture and exploit variable dependencies that standard GAs miss.
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 a Bayesian Genetic Algorithm when the optimization problem has strong dependencies (epistasis) among decision variables that defeat standard crossover, when fitness evaluations are expensive and sample efficiency matters, or when you need the algorithm to learn and exploit problem structure automatically. BGA excels on combinatorial problems with block structure (e.g., scheduling, feature selection, configuration optimization). Do NOT use it when the problem is separable (no variable interactions), when population sizes must be very small (the Bayesian network learner needs adequate samples), when real-time speed is critical (model learning adds overhead), or when a simple GA or gradient-based optimizer already performs well.
Strengths & limitations
- Captures and exploits variable dependencies through a learned Bayesian network, avoiding premature convergence caused by blind crossover.
- Highly sample-efficient on problems with strong epistasis, reaching good solutions in fewer function evaluations than standard GAs.
- The learned probabilistic model doubles as an interpretable summary of what high-quality solutions look like.
- Naturally handles discrete, continuous, and mixed variable types by choosing appropriate network parameterizations.
- Scales gracefully with problem structure: richer dependency graphs are learned only when the data support them.
- Learning a Bayesian network at every generation adds computational overhead; runtime per generation is higher than classic GA.
- Requires a sufficiently large population to learn reliable network structure; small populations yield poor model estimates.
- Network structure learning is NP-hard in general; practical implementations use heuristic search that may miss the true dependency graph.
- Performance degrades on problems with no meaningful variable interactions, where the modeling cost yields no benefit.
Frequently asked
How does a Bayesian Genetic Algorithm differ from a standard Genetic Algorithm?
A standard GA uses crossover and mutation operators that recombine solutions without regard to variable dependencies. A BGA replaces those operators with a Bayesian network learned from elite individuals, so offspring are sampled from a distribution that explicitly models how variables interact in high-quality solutions.
Is BGA the same as Bayesian optimization?
No. Bayesian optimization (BO) uses a surrogate model (e.g., Gaussian process) of the fitness landscape to pick the next single evaluation point via an acquisition function. BGA is an evolutionary algorithm that uses a Bayesian network to model the distribution of good solutions and generates an entire population of new candidates per generation.
What population size is recommended for BGA?
Empirically, populations of a few hundred to a few thousand are common. Pelikan et al. showed that the required population size scales polynomially with the number of decision variables for bounded-complexity Bayesian networks, so problems with many variables may need larger populations.
Can BGA handle continuous variables?
Yes. Continuous extensions use conditional Gaussian or mixed Bayesian networks (e.g., the EGNA or EMNA variants). Alternatively, continuous variables can be discretized, though this trades precision for modeling simplicity.
When does BGA fail to outperform a simple GA?
On separable or near-separable fitness landscapes (no meaningful variable interactions), BGA's model-learning overhead produces no benefit and a well-tuned standard GA will be faster and equally accurate.
Sources
- Pelikan, M., Goldberg, D. E., & Cantu-Paz, E. (1999). BOA: The Bayesian optimization algorithm. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-1999), pp. 525–532. Morgan Kaufmann. link ↗
- Larranaga, P., & Lozano, J. A. (Eds.) (2002). Estimation of Distribution Algorithms: A New Tool for Evolutionary Computation. Kluwer Academic Publishers, Boston. ISBN: 9781461352747
How to cite this page
ScholarGate. (2026, June 3). Bayesian Genetic Algorithm — Probabilistic model-guided evolutionary optimization. ScholarGate. https://scholargate.app/en/simulation/bayesian-genetic-algorithm
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 Multi-Objective OptimizationSimulation↔ compare
- Bayesian OptimizationOptimization↔ compare
- Genetic AlgorithmOptimization↔ compare
- Particle Swarm OptimizationOptimization↔ compare
- Stochastic Genetic AlgorithmSimulation↔ compare