Bayesian Optimization of Scenarios
Bayesian optimization tunes operating scenarios and design parameters by balancing exploration and exploitation over an expensive objective.
Optimizing an expensive black box
Many fusion design and scenario questions reduce to optimizing an objective that is expensive to evaluate and has no analytic gradient: maximize a performance metric over control settings, or over blanket and coil parameters. Bayesian optimization is built for exactly this regime of few, costly evaluations.
How it works
A probabilistic surrogate, often a Gaussian process, models the objective from the evaluations seen so far, including uncertainty. An acquisition function then proposes the next point to try by trading off exploiting regions that look good against exploring regions that are uncertain. The chosen point is evaluated and the surrogate updated.
- Surrogate: Gaussian process or Bayesian neural network with uncertainty
- Acquisition: expected improvement, upper confidence bound, or entropy search
- Loop: propose, evaluate, update, repeat until budget is exhausted
Constraints and multiple objectives
Real problems carry constraints, such as stability and hardware limits, and multiple goals at once. Constrained and multi-objective variants model feasibility and each objective separately, returning a set of good tradeoffs rather than a single optimum, which suits design where engineers weigh competing goals.
Strengths and limits
Bayesian optimization is sample-efficient and principled, but it scales poorly to very high-dimensional spaces and depends on a good surrogate. Dimensionality reduction and structured priors extend its reach. It complements, rather than replaces, brute-force scans where evaluations are cheap.
Applied to the Hyperion breeder concept, Bayesian optimization helps search toward the design point over parameters like current and shaping using surrogate objectives. The results are computational proposals for a machine under design, always confirmed with high-fidelity models before adoption.