An Example-Based Learning Approach to Multi-Objective Programming

In real-world optimization problems involving multiple objectives, the
weights of the objectives may not be specified, whereas example
solutions, i.e., the pairs of an instance and a solution, prepared by human
experts are usually available. This paper proposes a method for
determining the objective weights by using example solutions as the training set
so that a search algorithm can find reasonably good solutions for
all the instances. Our proposed method generates neighborhood
solutions defined by the search algorithm for each example solution,
and determines the weight settings. The method was successfully
applied to a scheduling problem in the steel manufacturing industry.