pabloo22 / gnn_scheduler Goto Github PK

The explanation should be contextualized for the JSP.

Part of the Phase 1 (Issue #4)

Rename package from jsspto job_shop and update imports

First, we need to define what we will do in this project. This is why a good literature review would be very beneficial before writing code. This literature review should be documented in the thesis and should answer the following questions:

• [Introduction?]: Why would anyone use ML to solve the JSP instead of other traditional methods such as MIP?
• [State of the art?]: What are the advantages of representing the problem as a graph instead of other modalities?
• [State of the art]: What is the best graph representation for the problem?
• [???]: What are the characteristics of the best GNN architectures in terms of representation and generalization?
• [???]: What are the pros and cons of using SL versus RL in this context?

Other things that should be explained in this phase (base knowledge):

• [Introduction?]: What are GNNs?
• What the Job Shop Scheduling Problem is about.
• What the role of dispatching rules is.
• Decide which experiments to conduct and when.

Deliverables

The following sections of the main body of the thesis should be written at the end of this phase:

• Introduction: Introduction to the problem (brief historical context), including the motivation and objectives of the Final Degree Project (TFG), as well as the structure of the report. This would be a first draft, it probably will be changed later on.
• State of the art: scope of the problem, its history, definitions, and related properties (the scientific and technological substrate).
• Technologies and tools to be used. Explanation of how PyG or other RL frameworks work (if we finally decide to use RL).
• Planning and cost of the work.

Full project deadline: May 15th

Phases

Phase 1: Literature Review

Phase deadline: February 15th
See Issue #4

Phase 2: Environment and Dataset creation

Phase deadline: March 15th

• Create a Dataset with solved instances.
• Create a Gymnasium Environment.

Phase 3: Neural Architecture Search and Pre-Training

Phase deadline: April 15th

• Create experiments for predicting similarity between operations based only on job and machine assignment.

Phase 4: Final experiments solving the JSP

• Change dtype of adj_matrix to torch.float32
• Change the shape of adj_matrix from [N, N, 2] to [2, N, N]