Doctoral thesis defense in Artificial Intelligence by Mr. Abebaw Degu

Euromed University of Fes (UEMF) is pleased to inform the public of the defense of the doctoral thesis in ‘‘Artificial Intelligence’’

The thesis defense will take place on Saturday, June 22, 2024 at 09:30 at UEMF
Location : the Main Hall of the Incubator (LOC001994) 

The thesis will be presented by Mr. Abebaw Degu WORKNEH under the theme:
”Synergizing Intelligence and Self-Improvement in Job Shop Scheduling for Smart Manufacturing”

Abstract

Intelligentization, supported by AI methodologies, has emerged over the last decades as a significant driver for manufacturing industries, propelling the growth of smart manufacturing. Classical AI has been endowed with greater capabilities in contemporary industries, resulting in industrial AI, now the technological foundation of the smart manufacturing industry. Smart manufacturing environments present unique challenges in job shop scheduling due to their complex machine environments, real-time data streams, and dynamic nature, necessitating dynamic and adaptive scheduling approaches to optimize production processes effectively. The complexity of smart manufacturing, where machines are interconnected and production environments change rapidly, makes it difficult for traditional scheduling techniques to be operational. RL’s adaptability and data-driven decision-making abilities provide a potentially effective solution. This thesis investigates the practical applications of reinforcement learning (RL) algorithms to address these challenges and optimize job-shop scheduling in the context of smart manufacturing. A DRL model that consists of different features and dynamic events, such as machine failure and job rework, is proposed for the Job Shop Scheduling Problem (JSSP) to minimize the makespan. The scheduling problem is formulated as a sequential decision-making process to visualize the interactive nature of the actual production environment. A general scheduling solution approach has also been developed for a Flexible Job Shop Scheduling Problem (FJSSP). The approach is a Triple Deep Q Network (TDQN) characterized by machine failure, job insertion, machine setup, and random processing time. The distinct feature of the approach is the ability to efficiently and consistently generate a schedule with lower total tardiness, its ability to learn optimal policy quickly (in fewer iterations) than DQN and DDQN, and its stability. The DRL models for JSSP are evaluated by applying them to well-known OR-tools dataset instances. On the other hand, the TDQN model is trained and evaluated on a hypothetically generated instance, which is, to our knowledge, the first

DRL approach and a scheduling solution with multiple dynamic events in an FJSSP environment, which is difficult to find. For problems with available benchmarks, it is indicated that the proposed DRL solution is competitive with the best method in the literature. Likewise, the TDQN approach’s result is evaluated based on the model’s performance in learning a stable policy and achieving minimal total tardiness compared to the DQN and DDQN models with a similar FJSSP configuration. In conclusion, this study underscores the main advantages of RL in improving job shop scheduling within smart manufacturing environments. By adopting RL, manufacturers can address real-world problems and maximize production efficiency, increasing their competitiveness in the smart manufacturing ecosystem.
 

Keywords— Deep Reinforcement Learning, Job Shop Scheduling, Smart Manufacturing, Intelligent Scheduling, Optimization, Sequential Decision Making

This thesis will be presented before the jury members: