A New Reward Function for Bayesian Feasibility Determination

Abstract

In Bayesian feasibility determination, a typical reward function is either the 0-1 or linear reward function. We propose a new type of reward function for Bayesian feasibility determination. Our proposed reward function emphasizes the importance of barely feasible/infeasible systems whose mean performance measures are close to the threshold. There are two main reasons why the barely feasible/infeasible systems are more important. First, the overall accuracy on solving a feasibility determination problem is heavily affected by those difficult systems. Second, if the decision maker wants to further find the best feasible system, it is likely that one of the barely feasible/infeasible systems is the best feasible. We derive a feasibility determination procedure with the new reward function in a Bayesian framework. Our experiments show that the Bayesian optimal procedure with the new reward function performs the best in making correct decisions on difficult systems when compared to existing procedures.