Make assumptions about function characteristics to evaluate possible test points. Pick optimal evaluation point.
Global optimization for each decision?
- No. Too expensive computationally.
Sampling decision space
- Limited computational expense
- Approximates optimal decision
In actuality, we avoid the computational complexity of dealing with complex probability distributions by encoding our assumptions in a function which ranks possible points to evaluate.
However, even so, there's still a tradeoff: The computational cost of optimizing our decision procedure can outweigh its benefit. We therefore limit our computational cost by sampling the decision space, thus approximating an optimal decision.