Bayesian Optimization for Experiments
Choosing the next expensive experiment or simulation to run so that few evaluations find a good setting.
The setting
Some fusion objectives can only be evaluated by an expensive experiment or simulation, and each evaluation is scarce. Bayesian optimization finds good inputs with as few evaluations as possible by building a probabilistic model of the objective and using it to pick each next point.
How it works
- Fit a surrogate, often a Gaussian process, to evaluations so far
- Use an acquisition function to score candidate next points
- Evaluate the highest-scoring point, then update the surrogate
- Repeat until the budget is spent or the objective plateaus
Exploration versus exploitation
The acquisition function balances trying points likely to be good (exploitation) against points where the surrogate is uncertain and might reveal something better (exploration). Common choices are expected improvement and upper confidence bound. This balance is why so few evaluations can suffice.
Uses in fusion
Bayesian optimization is applied to scenario tuning, actuator settings, and design parameter searches, wherever each trial is costly and the input space is moderate. It naturally respects that experiments are scarce, which is the fusion reality.
Limits
It works best in moderate dimensions and with a well-behaved objective; very high-dimensional or highly discontinuous problems are harder. Real experiments add noise and constraints (safety, hardware limits) that the method must respect. And the surrogate's assumptions matter: a poor kernel or model can steer the search astray. Used within these limits, it is an efficient way to spend a scarce evaluation budget.