Partially observable Markov decision process to generate policies in software defect management

Abstract

Bug repositories are dynamic in nature and as new bugs arrive, the old ones are closed. In a typical software project, bugs and their dependencies are reported manually and gradually using a issue tracking system. Thus, not all of the bugs in the system are available at any time, creating uncertainty in the dependency structure of the bugs. In this research, we propose to construct a dependency graph based on the reported dependency-blocking information in a issue tracking system. We use two graph metrics, depth and degree, to measure the extent of blocking bugs. Due to the uncertainty in the dependency structure, simply ordering bugs in the descending order of depth and/or degree may not be the best policy to prioritize bugs. Instead, we propose a Partially Observable Markov Decision Process model for sequential decision making and Partially Observable Monte Carlo Planning to identify the best policy for this sequential decision-making process. We validated our proposed approach by mining the data from two open source projects, and a commercial project. We compared our proposed framework with three baseline policies. The results on all datasets show that our proposed model significantly outperforms the other policies with respect to average discounted return.