Page 198 - IJOCTA-15-4

Z. Chen et al. / IJOCTA, Vol.15, No.4, pp.738-749 (2025)
search space. The main contributions of this work divided into three categories including cargo fea-
are as follows. tures, cargo hold features, and loading features.
• Optimization problem modeling. To
address the first challenge, FastLoader in-
troduces cargo loading constructor (§ 3.2) 2.2. Combinatorial optimization methods
that builds a data structure of cargo in- Combinatorial optimization is a mathematical op-
formation and complex constraints, which timization algorithm where each possible cargo
reduces the requirement of expertise in the position is explored under the simple constraints.
modeling process and provides data input The aim is to minimize center-of-gravity vari-
for the search space reduction stage. ation and satisfy all constraints. Existing re-
• Search space reduction. To address search can be divided into linear programming
the second challenge, FastLoader intro- such as COM, 6,19–22 nonlinear programming such
duces a search-space reduction stage that as IOM, 5,23,24 and mixed-integer linear program-
augments conventional heuristic optimiza- ming frameworks such as MLIP. 25–28
tion driven by LLM. Concretely, we con- COM 6 presents a joint integer linear pro-
struct a cargo-loading solver (§ 3.3), which gramming model that solves the discretization of
enumerates candidate load plans from the cargo under simple constraints. The model maxi-
raw cargo manifest and engineered feature mizes loading capacity and minimizes the center-
set. We also design a search space reducer of-gravity deviation. It performs well when con-
(§ 3.4), which scores the resulting solu- straints are simple but fails under complex con-
tions and prunes those that violate key straints. Thus, it cannot deal with the cargo load-
operational constraints extracted from the ing case investigated in this study. IOM pro-
5
LLM engine. poses an algorithm for the NP-hard cargo load-
Summary of experimental results. We ing problem. The algorithm splits the task into
evaluate FastLoader on four common aircraft four sub-problems: aircraft configuration, pallet
types, and we conduct the evaluation using seven assembly, air-cargo palletization, and center-of-
baselines across five LLMs. The results show the gravity balancing. It integrates path optimiza-
following: (1) FastLoader achieves the best per- tion to maximize transport benefit. However, the
formance compared to the seven baselines with four sub-problems do not cover all constraints
only 1.5% accuracy loss; (2) in the comparison of in our scenario, so the method performs poorly
time consuming between different heuristic search when constraints become complex. MLIP 25 intro-
baselines and FastLoader, FastLoader reduces the duces a mixed-integer programming model that
time costs by 90% on average; (3) FastLoader optimizes cargo center-of-gravity and inertia mo-
achieves stable performance across 5 LLMs, which ment to improve loading efficiency and reduce fuel
proves FastLoader is applicable to multiple LLMs. burn. Using real data constraints, the method
greatly shortens search time. However, its data
2. Background & related work
modeling fails to cope with highly constrained
In the cargo loading-optimization scenario(§ 2.1), scenes. When constraints increase, MLIP cannot
existing methods can be divided into the always find feasible solutions. MLIP-ACLPDD 28
following two categories: (1) combinatorial employs a similar mixed-integer model. Its ob-
optimization(§ 2.2) methods are designed to solve jective minimizes fuel consumption and loading
cargo loading optimization under simplified con- cost. The method still shows low accuracy under
straints; (2) heuristic search (§ 2.3) methods are complex constraints.
designed to solve cargo loading optimization un- In conclusion, when applied to complex con-
der complex constraints. straints scenario, existing combinatorial optimiza-
tion methods lead to the following two main draw-
2.1. Background backs: (1) complex constraint scenario requires
In typical air cargo loading scenario, cargo people with professional knowledge to model, and
types fall into two broad categories: Unit Load the scenario modeling is difficult; (2) combinato-
Devices (ULDs) are large cargo containers suit- rial optimization methods have difficulty finding
able for wide-body aircraft (e.g., B777, B787). the optimal solution within the modeling of com-
Bulk cargo comprises loose cartons, mailbags, plex constraints. As a result, the accuracy of com-
live animals, and other items placed directly into binatorial optimization methods is limited by the
the bin-shaped holds of narrow-body aircraft(e.g., complex constraints scenario. Heuristic search is
B737, A320). As shown in Table 1, there are 17 designed to solve these two drawbacks when the
complex constraints in our scenario, and they are constraints become complex.
740