A constraint-aware Bayesian optimization policy selects evaluations by balancing objective quality with practical feasibility. It targets the best feasible point rather than the unconstrained global optimum, which may be too costly, dangerous, or impractical to use. section 8.0 section 8.6
The method proceeds by modeling constraints and incorporating them into acquisition decisions:
- Classify points that satisfy every constraint as feasible and points that violate one or more constraints as infeasible. section 8.6
- Train a Gaussian process on each function defining a constraint, alongside the Gaussian process for the objective. section 8.6
- Estimate the probability that a candidate is feasible from the constraint GPs’ normal predictive distributions. section 8.6
- Modify expected improvement by incorporating feasibility probability into its acquisition score, balancing objective improvement against constraint satisfaction. section 8.6
Neural-network architecture tuning
Increasing network depth may improve accuracy while making training or deployment too expensive. A constraint-aware policy therefore searches for high-accuracy architectures that remain computationally feasible instead of pursuing the unconstrained accuracy maximum. section 8.0
