A dynamic learning-based genetic algorithm for scheduling resource-constrained projects with alternative subgraphs

Genetic algorithms (GAs) are population-based algorithms widely applied for solving complex scheduling problems and such the resource-constrained project scheduling problem with alternative subgraphs (RCPSP-AS) in which alternatives for work packages should be selected prior to project scheduling. The objective of this research is twofold. First, we develop a dynamic GA based on a new hybridisation of initialisation procedures, local searches based on learning approaches and restart schemes for scheduling problems in general. Second, we improve existing benchmark solutions for a large artificial dataset for the RCPSP-AS in particular. Our dynamic GA leverages existing constructive heuristics and priority rules to create a pool of high-quality initial solutions. Subsequently, these solutions are further improved by means of learning approaches that are designed as weight- or population-based local searches. In order to avoid getting stuck in a local optimum, various restart schemes are implemented. Based on our results, gradual learning and learning based on the population outperform other approaches for high-complex problem instances. Since metaheuristics — such as GAs — are mainly beneficial in complex problem settings, we are convinced that these research findings can inspire researcher when solving similar or other scheduling problems.