- Active learning: estimate an accurate model
- Start with prior function (can assume uniform)
- Evaluate point with greatest uncertainty (usually highest variance and furthest away from last evaluated point)
- Go back to 1 until model is accurate enough or budget runs out
- Bayesian optimization: find global optimization
- Start with prior function (can assume uniform)
- Evaluate next point by optimizing an acquisition function around our current knowledge
- Probability of improvement
- Expected improvement
- Based on prior knowledge, where should we evaluate next?
- Active: most uncertain
- Bayesian: balance between uncertain and current optimums