Feature Dimensions in CVRPTW/MTVRP Encoder

[Ynantao]

Subject: Feature Dimensions in CVRPTW/MTVRP Encoder

I have been using the code provided for solving CVRPTW and MTVRP problems and encountered a possible issue regarding the feature dimensions in the encoder part of the model.

In the CVRPTWModel and MTVRPModel, both models seem to use 5 features for encoding in the node_xy_demand_TW tensor:

• x_coord
• y_coord
• demand
• earlyTW (early time window)
• lateTW (late time window)

However, in the MTVRPModel encoder, the comment mentions that 6 features are being used. Specifically, it says:

# 6 features are: x_coord, y_coord, demands, earlyTW, lateTW

But only 5 features are listed. The input tensor node_xy_demand_TW is still 5-dimensional as in the CVRPTWModel.

Could you kindly clarify whether this is:

1. A typo in the comment, and the correct number of features is 5?
2. Or is the model intended to have an additional (6th) feature that is missing from the input processing (e.g., task type, service time, or another relevant feature for MTVRP)?

[FeiLiu36]

Thank you for your interest in the work.

Sorry for the late reply. I have just checked the code and found it is a typo in the comment. It has been revised.

Moreover, I think specifying the type within the encoder might be helpful to improve its representation as a future improvement.