Page 70 - Read Online

Page 10 of 15 Ma et al. Intell Robot 2023;3(4):581-95 I http://dx.doi.org/10.20517/ir.2023.33

in the ACO algorithms are comprehensively updated. This dual optimization process enables graph feature
learning and path search to promote each other and gradually improve the effectiveness of graph data process-
ing. After multiple iterations, when the algorithm meets the convergence conditions, GHNN-ACO outputs
the final graph features and node classification results.

4. RESULTS
In this study, we explore the application of GHNN-ACO algorithms in real-world scenarios involving monitor-
ingandrespondingtoemergencies. Bytestingthealgorithminthesescenarios, wehaveconductedanin-depth
analysis of its performance in node classification and link prediction tasks within a multi-agent system.

4.1. Parameters setting and experimental subjects
In this design, we set up an emergency rescue task environment to simulate the emergency economic rescue
tasks after a natural disaster occurs. Assuming that the affected area is represented by a 10x10 grid map, with
each grid representing a task location. There are three different types of tasks to be carried out: searching and
rescuing disaster victims, transferring affected residents, and distributing rescue supplies. The priority of tasks
is divided into three levels: high, medium, and low.

In the task environment, we define the following parameters:

(1) Map size: 10 × 10 grid map with a total of 100 task locations.

(2) Task quantity: There are a total of ten tasks to be executed, including three tasks for searching and rescuing
disaster victims, three tasks for transferring disaster-affected residents, and four tasks for distributing rescue
supplies.

(3) Task type: Search and rescue disaster victims, transfer disaster-affected residents, and distribute rescue
supplies.

(4) Task Priority: Each task is assigned one of three priority levels: high, medium, or low.

(5) The experimental settings and initial conditions are as follows: Drone count: 2; Unmanned ship count: 1;
Unmannedvehiclecount: 2. Intheinitialstate,allagentsandtasksarerandomlypositionedandassignedstates.
The intelligent agents can move freely on the map but cannot occupy the same task location simultaneously.
The priorities and types of tasks are also randomly assigned.

4.2. Experimental test results
OurresearchresultsindicatethattheGHNN-ACOalgorithmperformswellinprocessingheterogeneousmulti-
agent data. By utilizing GHNNs, this algorithm efficiently handles different types of agents and relationships in
intelligent agent systems. The ability to process heterogeneous graph data enables the GHNN-ACO algorithm
to better capture complex associations and features between intelligent agents, thereby effectively improving
the accuracy of node classification and link prediction tasks.

The algorithm used in this paper can quickly allocate different tasks to the required executing agents in five
task scenarios. Figure 3(A) shows the task allocation between intelligent vehicles and UAVs. It can be seen
that the algorithm in this article can autonomously allocate tasks to the first two scenarios, while the allocation
efficiency for the third scenario involving unmanned ships is not as effective. Figure 3(B) and 3(C) represents
tasks that focus on unmanned vehicles and unmanned ships and emergency rescue tasks that involve UAVs
and unmanned ships, respectively.