Page 205 - IJOCTA-15-4

FastLoader: Leveraging large language models to accelerate cargo loading optimization with numerous
the inference does not exceed 3 s, indicating that reduction stage, the search space reducer effec-
in different large model deployments, the opti- tively reduces the search time and improves the
mization ability of FastLoader is robust under dif- solution speed. It achieves a significantly shorter
ferent population size settings. runtime while sacrificing only a negligible amount
Other parameter setting analysis. As of solution accuracy.
shown in Figure 5j-l, we evaluate four differ- We are further improving search efficiency in
ent combinations of temperature, top p and re- the solution process for cargo hold loading prob-
peat penalty. The four parameter combinations lems. This aims to handle more complex and di-
are: {0.8,0.9,1.2}, {0.3,0.3,1.0}, {0.6,0.7,1.5}, and verse loading scenarios effectively. Complex sce-
{0.5,0.6,1.1}. We have the following two key ob- narios often involve more constraints across mul-
servations: (1) in terms of algorithm runtime, the tiple dimensions. Examples include item compat-
runtime of Deepseek and Llama improves in the ibility, stacking stability, and loading/unloading
first three sets of parameter settings. However, for order. To obtain high-quality solutions in a
the other LLMs, FastLoader still has stable run- shorter time, we aim to integrate reinforcement
time under multiple experimental combinations learning into the search space reduction process
and still has good robustness; (2) in the first three in the future.
sets of parameter settings, the solution accuracy
of LLM remains unchanged and has high robust- Acknowledgments
ness. But the accuracy decreases slightly under None.
the fourth parameter combination. This is be-
cause under this parameter setting, temperature Funding
is 0.5, which introduces moderate generation di- This work was supported by the National Nat-
versity, while top p is 0.6, which limits the selec-
ural Science Foundation of China (Grant No.
tion space. They cause the model generation to
62272046, 62132019, 61872337), the high-quality
deviate slightly from the optimal solution, but the
development project of the Ministry of Industry
accuracy remains at a high level.
and Information Technology (CEIEC-20240), Sci-
Impact of Model Stability on Perfor-
ence and Technology Innovation Project of Bei-
mance. We further analyze the impact of large jing Institute of Technology (2023CX01017), and
model stability on our method’s performance. In Open Project of Key Laboratory of the Ministry
Stage one, modeling relies on prompts to help of Education (2023PY002).
the model understand complex scenarios. Fig-
ure 5 d-f, j,k reflects this effect. When the length Conflict of interest
of tokens increases and the repetition penalty is
too high, inference time rises and accuracy drops. The authors declare they have no competing in-
This is because a high repetition penalty makes terests.
the model’s judgment unstable in complex con- Author contributions
strained settings. As a result, it fails to reduce
the search space efficiently, lowering both speed Conceptualization: Zheng Chen, Chi Harold Liu,
and solution quality. Rui Han
Results. Across iterations, token lengths, Investigation: Yunlai Cheng, Chi Harold Liu,
population sizes, and other parameter settings, Yinglong Wang
FastLoader maintains high solution accuracy Methodology: Yunlai Cheng, Zhe Ouyang , Jing
within 3 s, confirming its robustness while expos- Chen
ing less than 1% differences among the evaluated Formal analysis: Yunlai Cheng, Xinran Li , Yue
LLMs. Han
Writing – original draft: Yunlai Cheng, Zheng
Chen, Rui Han
5. Conclusion Writing – review & editing: Siqi Du, Ying Guo ,
Dongzhou Zhao, Meng Yu
This work presents the design and evaluation
of FastLoader, an LLM-based optimization ap- Availability of data
proach for accelerating cargo loading. In the op-
timization problem modeling stage, cargo loading Not applicable.
constructor accurately model the scenarios with
AI tools statement
numerous cargo types and complex constraints. It
achieves higher solution accuracy than combina- All authors confirm that no AI tools were used in
torial optimization methods. In the search space the preparation of this manuscript.
747