Learn from the Past: Improving Model-Directed Optimization by Transfer Learning Based on Distance-Based Bias

For many optimization problems it is possible to define a problem-specific distance metric over decision variables that correlates with the strength of interactions between the variables. Examples of such problems include additively decomposable functions, facility location problems, and atomic cluster optimization. However, the use of such a metric for enhancing efficiency of optimization techniques is often not straightforward. This paper describes a framework that allows optimization practitioners to improve efficiency of model-directed optimization techniques by combining such a distance metric with information mined from previous optimization runs on similar problems. The framework is demonstrated and empirically evaluated in the context of the hierarchical Bayesian optimization algorithm (hBOA). Experimental results provide strong empirical evidence that the proposed approach provides significant speedups and that it can be effectively combined with other efficiency enhancements. The paper demonstrates how straightforward it is to adapt the proposed framework to other model-directed optimization techniques by presenting several examples.