incumbent
The current best value used to define the baseline for improvement, serving as the comparison benchmark for acquisition functions in Bayesian optimization and simulation optimization. Its selection affects the balance between exploration and exploitation.
Timeline (last 90 days)
The three core design choices of the EI algorithm are: the incumbent that defines the improvement benchmark, the number of replications per sampling point (replication level), and the terminal recommendation.
The paper finds that when minimizing the regret upper bound, the incumbent should choose a lower quantile coefficient, and this coefficient exactly matches the coefficient in the confidence bound of the SK predictor.
The paper proposes the design philosophy of 'optimistic in search, conservative in delivery': the intermediate simulation steps adopt an optimistic incumbent to search sufficiently, while the terminal recommendation remains conservative with respect to the remaining posterior uncertainty.
All Facts (3)
The three core design choices of the EI algorithm are: the incumbent that defines the improvement benchmark, the number of replications per sampling point (replication level), and the terminal recommendation.
50%UnverifiedThe paper finds that when minimizing the regret upper bound, the incumbent should choose a lower quantile coefficient, and this coefficient exactly matches the coefficient in the confidence bound of the SK predictor.
50%UnverifiedThe paper proposes the design philosophy of 'optimistic in search, conservative in delivery': the intermediate simulation steps adopt an optimistic incumbent to search sufficiently, while the terminal recommendation remains conservative with respect to the remaining posterior uncertainty.
50%