Multi-objective Integer Programming Approaches to Next Release Problem - Enhancing Exact Methods for Finding Whole Pareto Front
Shi Dong,Yinxing Xue,Sjaak Brinkkemper,Yan-Fu Li
DOI: https://doi.org/10.1016/j.infsof.2022.106825
IF: 3.9
2022-01-01
Information and Software Technology
Abstract:Context: Project planning is a crucial part of software engineering, it involves selecting requirements to develop for the next release. How to make a good release plan is an optimization problem to maximize the goal of revenue under the condition of cost, time, or other aspects, namely Next Release Problem (NRP). Genetic and exact algorithms are used since it was proposed. Objective: We model NRP as bi-objective (revenue, cost) and tri-objective (revenue, cost, urgency) form, and investigate whether exact methods could solve bi-objective and tri-objective instances more efficiently. Methods: The state-of-art integer linear programming (ILP) approach to the bi-objective NRP is epsilon-constraint for finding all non-dominate solutions. To improve its efficiency, we employ CWMOIP (Constrained Weighted Multi-Objective Integer Programming) and I-EC (improved epsilon-constraint) for solving bi-objective instances. In tri-objective form, we introduce SolRep, an ILP method that optimizes the reference points from sampling, for finding solutions subset within a short time. NSGA-II is implemented as the evolutionary algorithm for the comparison with former methods and it adopts the seeding mechanism. Results : I-EC can find all non-dominated solutions with better performance than both epsilon-constraint and CWMOIP on all instances except for one. I-EC reduces solving time by 19.7% (large instances) and 91.5% (small instances) on average separately compared with epsilon-constraint. SolRep can find evenly distributed solutions and exceed NSGA-II illustrated by several indicators (such as HyperVolume) on tri-objective instances. And each method has its merit in the aspect of speed and number of the solutions. Conclusion: (1) The I-EC can solve all non-dominated solutions with better performance than the state-of-art exact method. (2) SolRep solves large tri-objective instances with more non-dominated solutions and solves small instances with less time compared with seeded NSGA-II. (3) Seeded NSGA-II shows its advantage on the number of non-dominated solutions on smaller tri-objective instances.