Multi-workflow dynamic scheduling in product design: A generalizable approach based on meta-reinforcement learning

Abstract

Multi-stage simulation workflows are a commonly used and crucial means for evaluating and optimizing design performance throughout the stages of product development. Effective multi-workflow scheduling is crucial for ensuring optimal resource utilization and execution time. Addressing simulation multi-workflow dynamic scheduling (SWDS) is challenging due to the variability of tasks and the uncertainty in execution, necessitating flexible and adaptive scheduling strategies. While traditional methods such as heuristic-based algorithms are popular in workflow scheduling, they show weaknesses in robustness, generalization, and adaptability when dealing with highly dynamic environments such as SWDS. To address above issues in SWDS, this paper innovatively proposes a meta-reinforcement learning-based scheduling method that aims to enhance generalization and adaptability to dynamic conditions. An enhanced Model-Agnostic Meta-Learning based deep reinforcement learning (DRL) algorithm is proposed to acquire dynamic scheduling strategies through multi-scenario training. Multi-step state features are extracted to address the issue of insufficient state observations. Conjugate adaptive search and Armijo conditions are applied to enhance the effectiveness of algorithm training. Experimental tests in 180 multi-type scenarios, compared with nine heuristic methods and three state-of-the-art DRL algorithms, comprehensively demonstrate the superiority of the proposed method.