An Investigation of Terminal Settings on Multitask Multi-objective Dynamic Flexible Job Shop Scheduling with Genetic Programming

Multitask learning has attracted widespread attention to handle multiple tasks simultaneously. Multitask genetic programming has been successfully used to learn scheduling heuristics for multiple multi-objective dynamic flexible job shop scheduling tasks simultaneously. With genetic programming, the learned scheduling heuristics consist of terminals that are extracted from the features of specific tasks. However, how to set proper terminals with multiple tasks still needs to be investigated. This paper has investigated the effectiveness of three strategies for this purpose, i.e., intersection strategy to use the common terminals between tasks, separation strategy to apply different terminals for different tasks, and union strategy to utilise all the terminals needed for all tasks. The results show that the union strategy which gives tasks the terminals needed by all tasks performs the best. In addition, we find that the learned routing/sequencing rule by the developed algorithm with union strategy in one multitask scenario can share knowledge between each other. On the other hand and more importantly, the learned routing/sequencing rule can also be specific to their tasks with distinguished knowledge represented by genetic materials.