Learning the Quality of Dispatch Heuristics Generated by Automated Programming

摘要

One of the challenges within the area of optimisation, and AI in general, is to be able to support the automated creation of the heuristics that are often needed within effective algorithms. Such an example of automated programming may be performed by search within a space of heuristics that will be applied to a target domain. In this, brief proof-of-concept, paper, we consider the case of online bin-packing as the target domain, and consider the potential for machine learning methods to aid the associated automated programming problem. Simple numerical 'policy matrices' are used to represent heuristics, or 'dispatch policies', controlling the placement of item into bins as they arrive. We report on an initial investigation of the potential for neural nets to analyse and classify the resulting 'policy matrices', and find strong evidence that simple nets can be trained to learn to predict which heuristics, expressed as policy matrices, exhibit better or worse fitness. This gives the potential for them to be used as a surrogate fitness function to enhance the usage of search algorithms for finding heuristics. It also supports the prospect of using machine learning to extract the patterns that lead to successful heuristics, and so generate explanations and understanding of machine-generated heuristics.