Page 202 - IJOCTA-15-4

Z. Chen et al. / IJOCTA, Vol.15, No.4, pp.738-749 (2025)
Qwen-1.5-72B 42 is Alibaba’s 72-billion-parameter loading, it is calculated as
multilingual model that supports up to 32 k con- W · (BA − RA)
text tokens and excels at chat, code, and tool Index = C + K (6)
use; (5) Qwen-3-235B 43 is a 235-billion-parameter
where BA is the horizontal distance in
mixture-of-experts model that activates roughly
units of length from the reference datum
22 billion parameters per token to achieve high
to the location. The smaller index change
accuracy with lower compute overhead.
rate represents the better performance of
Dataset. The experiments rely on an air algorithms.
cargo loading dataset collected by China Aviation • Algorithm runtime: Algorithm run-
Information Co. from June 2024 to January 2025 time refers to the measure of execution
at an airport in Beijing, China. The data were time. The lower algorithm runtime rep-
collected at the flight level, with each record corre- resents the better performance of algo-
sponding to a specific flight’s cargo loading infor- rithms.
mation. After initial cleaning to remove erroneous
entries, the dataset comprises over 6.1 million 4.2. Evaluation of loading performance
cargo records spanning more than 1000 aircraft Hyperparameter setting. For the genetic al-
types. For this study, we sample 200,000 records gorithms (HGA and GA-normal), we set a pop-
drawn from two wide-body aircraft (B777, B787) ulation of 10,000, run 100 generations, and use a
and two narrow-body aircraft (A320, B737).
mutation rate of 0.7 with a crossover rate of 0.2.
Baseline. We evaluate four state-of-the- For the particle-swarm optimizations (DMOPSO
art heuristic search methods, including HGA, 10 and PSO-normal), we employ 2000 particles and
9
8
GA-normal, 10 DMOPSO, and PSO-normal and 100 iterations. Our own method runs for 100 it-
three combinatorial optimization methods, in- erations on a search space of 100 individuals and
6
5
cluding COM, IOM and MLIP. 25 sets the LLM temperature at 0.8, top p 0.9, and
Metric Following the setting of, 44 we utilize a repetition-penalty of 1.2.
the following three metrics: Comparison of loading accuracy. Table 2
• %MAC: The Mean Aerodynamic Chord and Table 3 show the comparison results of all
(MAC) represents the chord of a hypo- eight methods and five testbeds. We have three
thetical rectangular wing whose projected key observations, which are following:
area equals that of the real wing, thereby (1) Generalization analysis of FastLoader.
maintaining the same aerodynamic and Compared with the combinatorial op-
moment characteristics. Although exact timization method, FasterLoader shows
computation is challenging, the MAC is higher and more stable solution accuracy,
often established via wind tunnel testing. with higher generalization performance.
As a key factor in aircraft load analysis, This is because the LLM fully under-
the %MAC is derived based on this mean stands the complex constraints in the air
aerodynamic chord. The MAC is calcu- cargo loading scenario. However, com-
lated as binatorial optimization methods are dif-
ficult to fully model complex constraints,
C·(I−K) + RA − LEMAC resulting in low solution accuracy.
%MAC = W × 100 (5) (2) Analysis of different large models. By
MAC
observing the accuracy results of the
where C is the weight constant of the air- five LLMs, we find that the accuracy of
craft, I is the index value corresponding Deepseek is lower than that of other large
to the respective weight, K is the constant models under different LLMs. This is be-
that is used to avoid the negative figures, cause Deepseek lacks the knowledge of the
W is the actual weight of the aircraft, relevant scenarios.
RA is the reference index values, LEMAC (3) Heuristic search algorithm analysis. We
is the horizontal distance in length units observe that without considering the al-
from reference datum to the location from gorithm runtime, the results of heuristic
the leading edge, and MAC is the length of search achieve the highest solution accu-
mean aerodynamic chord in length units. racy in the evaluation of all aircraft types.
The smaller %MAC represents the better This is because the heuristic search meth-
performance of algorithms. ods generate a huge search space to make
• The change rate of index: Index is an- the solution meet the complex constraints
other significant concept in aircraft cargo and improve the solution accuracy.
744