Master's Thesis

Reinforcement Learning for Job Shop Scheduling Problems

Ricardo Miguel Alves Magalhães — 2021

Key information

Authors:

Ricardo Miguel Alves Magalhães (Ricardo Miguel Alves Magalhães)

Supervisors:

Miguel de Sousa Esteves Martins (Miguel De Sousa Esteves Martins); Susana Margarida da Silva Vieira (Susana Margarida da Silva Vieira)

Published in

January 22, 2021

Abstract

There is a revolution going on in the way industries work. The large abundance of data and the exponential growth of computational power have unchained Artificial Intelligence’s potential. This new powerful tool gained popularity amidst the advent of Industry 4.0 and comes with the promise of making factories smart, more productive and efficient. The method proposed in this thesis applies Deep Q-Learning to create an intelligent system capable of solving the Dual Resource Constrained Flexible Job Shop Scheduling Problem. A novel Encoder-Decoder neural network architecture with Attention Mechanism is presented. This architecture allows the network to be fed more thorough and meaningful problem information, as well as having more precise and powerful output commands. In order to diminish the computational cost of this technique, the output commands are selected as in Pointer Networks architectures. The agent developed achieved promising results, even having just been scarcely trained. These results show that, with an optimized implementation of the method and a more thorough training procedure, this technique holds the potential of challenging current state-of-the-art meta-heuristic algorithms in solving this kind of problems. Finally, future work might consist in adding additional constraints to the problem for application in the Industry. The possibility of having different processing times for the same operation, depending on the allocated resource pair, is discussed. Also, a redesigned concept of the proposed method is presented.

Publication details

Authors in the community:

Supervisors of this institution:

Fields of Science and Technology (FOS)

mechanical-engineering - Mechanical engineering

Publication language (ISO code)

por - Portuguese

Rights type:

Embargo lifted

Date available:

December 3, 2021

Institution name

Instituto Superior Técnico