Benchmark
Results appear after an experiment.
Current: —
Assign and order jobs on identical parallel machines by encoding exactly m−1 machine separators inside one permutation. Compare total tardiness, maximum tardiness, makespan, completion time and load balance.
Results appear after an experiment.
Current: —
Parallel-machine tardiness is more than load balancing because order within each machine changes completion times and job tardiness Tⱼ=max(0,Cⱼ−dⱼ). Minimizing their sum jointly decides assignment and sequencing.
Separator encoding reuses the same permutation engines: separators are learnable nodes with positions and adjacencies. Distinct separator labels are semantically equivalent, however, creating a symmetry that should be reported in research.
The chromosome has n+m−1 unique tokens and always contains m−1 separators, decoding without repair into m machine sequences. Adjacent separators permit an empty machine, penalized naturally through makespan and load variance.
Let pⱼ be processing time, dⱼ due date, Cⱼ completion time on its assigned machine, and Lₖ machine load.
The single-objective track follows the classical total-tardiness objective. The MO track exposes conflicts among tardiness, throughput and balance, using nondominated sorting and Flow-Shop-equivalent convergence and Pareto reporting.
Tanaka and Araki publish identical-parallel-machine benchmarks for 2–10 machines with reported optimum results. The original archive is hosted on Google Drive and is not copied into this deployment; current preload entries are therefore disclosed as controlled fixtures using benchmark dimensions, not published instances. The archive can later replace fixtures without changing the evaluator or API.