Combining Multiple Optimization Runs with Optimal Dependency Trees

摘要

When trying to solve a combinatorial optimization problem, often multiplealgorithms and/or multiple runs of the same algorithm are used in order to find multiple local minima. We present a method which uses information from previous runs to determine promising starting points for future searches. Our algorithm, termed COMIT, models inter-parameter dependencies present in the previously found high-evaluation solutions. COMIT incrementally learns optimal dependency trees that model the pairwise dependencies in a set of good solutions found in previous searches. COMIT then samples the probability distributions modeled by these trees to generate new starting points for future searches. This algorithm has been successfully applied to jobshop scheduling, traveling salesman, knapsack, rectangle packing, and bin-packing problems.