This paper studies the feasibility of efficiently combining both a software component model and a task-based model. Task based models are known to enable efficient executions on recent HPC computing nodes while component models ease the separation of concerns of application and thus improve their modularity and adaptability. This paper describes a prototype version of the COMET programming model combining concepts of task-based and component models, and a preliminary version of the COMET runtime built on top of StarPU and L2C. Evaluations of the approach have been conducted on a real-world use-case analysis of a subpart of the production application GYSELA. Results show that the approach is feasible and that it enables easy composition of independent software codes without introducing overheads. Performance results are equivalent to those obtained with a plain OpenMP based implementation.
8 May 2023 in The International FLAIRS Conference Proceedings
Genetic algorithms (GAs) belong to the category of evolutionary algorithms and are frequently utilized for resolving challenging combinatorial problems. However, they typically require customization to suit a particular problem type, and their performance is heavily influenced by numerous hyperparameters and reproduction operators. In this work, we propose a Reinforcement Learning approach for fine-tuning Genetic Algorithms in Flexible Job Shop Scheduling problems (FJSP), where the main parameters involved in the genetic algorithm operators are trained to allocate the most promising values. The approach returns an optimized schedule taking into account given constraints specific to the scenario, such as the relationship among release date, due date, and processing time, which machines must be selected out of a set of alternative machines, or which sequence-dependent setup time can be filtered. The approach takes input data in the form of FJSP instances by varying the numbers of jobs and machines and then uses the NSGA-II algorithm to generate solutions. These solutions are stored in a Solutions module and they are analyzed using a Principal components analysis (PCA) to identify clusters of similar instances and solutions. The Q-Learning module then generates hyperparameters for each iteration of the NSGA-II algorithm based on information from the previous modules. A toy example is presented to better understand the behavior of the proposal and the results obtained for optimizing further instances of the problem in a more efficient way.