Using the k-Nearest Problems for Adaptive Multicriteria Planning

Abstract

This paper concerns the design and development of an adaptive planner that is able to adjust its parameters to the characteristics of a given problem and to the priorities set by the user concerning plan length and planning time. This is accomplished through the implementation of the k nearest neighbor machine learning algorithm on top of a highly adjustable planner, called HAP. Learning data are produced by running HAP offline on several problems from multiple domains using all value combinations of its parameters. When the adaptive planner HAP$_{\textrm{NN}}$ is faced with a new problem, it locates the k nearest problems, using a set of measurable problem characteristics, retrieves the performance data for all parameter configurations on these problems and performs a multicriteria combination, with user-specified weights for plan length and planning time. Based on this combination, the configuration with the best performance is then used in order to solve the new problem. Comparative experiments with the statistically best static configurations of the planner show that HAP$_{\textrm{NN}}$ manages to adapt successfully to unseen problems, leading to an increased planning performance.