FastLoader: Leveraging large language models to accelerate cargo loading optimization with numerous loading constraints

With the unquestionable commercial success of air cargo transportation, cargo loading is a crucial step that selects the optimal placement solution for a given aircraft hold and a set of cargoes. This combinatorial optimization promotes airlines’ revenue (e.g., minimizing fuel consumption) with the encoded constraints in the solution space. In practical scenarios, cargo loading includes dozens of loading constraints (e.g., isolation of dangerous cargoes). However, existing techniques either over-simplify such constraints due to the expensive manual modeling in combinatorial optimization, or suffer from a time-consuming optimization process due to the large search space in heuristic search. In this paper, we present FastLoader, an optimization acceleration approach that employs large language models (LLMs) to distinguish critical structural patterns in the simulated cargo loading data while still scaling to numerous loading constraints in real scenarios. FastLoader’s key design features are the following: (i) a cargo loading constructor, which converts the information of both cargo types and loading constraints into pre-defined data structures, thus avoiding manual modeling and improving solution accuracy; (ii) a cargo loading solver and a search space reducer, which work together to effectively reduce search space and accelerate the optimization process. We evaluate the proposed approach using a list of practical scenarios from industry transportation systems, and the results show the followin: FastLoader improves accuracy by 10% compared to combinatorial optimization, and reduces the optimization time by 90% with 1.5% accuracy losses compared to heuristic search.