L2S

L2S proposes a DRL-based improvement heuristic for JSSP. It encodes complete solutions as disjunctive graphs, leverages a dual-module GNN (TPM for topology, CAM for job–machine semantics), and trains a policy via n-step REINFORCE to generate swap operations without full neighborhood evaluation.

Usage

You can run the following command lines to execute the code.

python train.py/eval.py settings=l2s_settings mode={train/test} problem=jssp num_job={num_job} num_machine={num_machine} 
decoder_strategy={sampling/greedy} 

Tasks

Supported Tasks: JSSP.

Required Data Generator:

• JSSPGenerator

Training

L2s applies autoregressive reinforcement learning class ARREINFORCELightning() for training.