Page 62 - Read Online

Page 2 of 15 Ma et al. Intell Robot 2023;3(4):581-95 I http://dx.doi.org/10.20517/ir.2023.33

1. INTRODUCTION
Multi-agent systems have been widely used in various fields, such as robot cooperation, responding to intelli-
gent city emergencies, and distributed sensor networks. The multi-agent system in this paper is composed of
unmanned aerial vehicles (UAVs), unmanned ships, and unmanned vehicles, which work together efficiently
[1]
to respond to emergencies .

In these systems, multiple agents collaborate to complete tasks, thereby improving the efficiency and perfor-
mance of the system. As a key component of the multi-agent system, task allocation directly affects the overall
performance of the system [2,3] . However, due to the heterogeneity among agents and the dynamic nature of
tasks, the assignment problem becomes complex and challenging. Therefore, we need an efficient method to
solve the assignment problem of heterogeneous multi-agents [4,5] .

Graph heterogeneous neural network ant colony optimization (GHNN-ACO) algorithms are a new method
thatcombinesgraphheterogeneousneuralnetworks(GHNN)withantcolonyoptimization(ACO)algorithms
to solve the assignment problem of heterogeneous multi-agents [6,7] . In this method, the heterogeneous multi-
agent system is modeled as a graph structure, where the agents are represented as nodes within the graph.
The communication and collaboration relationships between agents are denoted as edges in the graph [8,9] .
Through the joint optimization of ACO algorithms and the neural network, this method can achieve efficient
and accurate task allocation, thus improving the performance of the entire multi-agent system [10] .

In a heterogeneous Multi-agent system, the heterogeneity of agents is reflected in their characteristics, capa-
bilities, and task requirements [11–14] . To account for this heterogeneity, we model the Multi-agent system as
a graph, where each agent is represented as a node, and the connection edges between nodes represent the
communication and collaboration relationships between agents. This graph structure accurately reflects the
relationships and interactions between intelligent agents [15–17] . ACO algorithms are optimization algorithms
that simulate the behavior of an ant colony when searching for the optimal path. In the task allocation of graph
heterogeneous multi-agent systems, we use ACO algorithms to simulate the process of information transmis-
sion and cooperation between agents. The ACO algorithms have the ability for global search, enabling agents
to quickly find the preliminary solution for task assignment. Their basic idea is that ants release pheromones
whensearchingforthepathandchoosethenextmovingdirectionbasedonthepheromoneconcentration [8,18] .
Agents simulate the cooperation process by releasing pheromones in the task allocation process and selecting
task allocation strategies according to the pheromone concentration [19–22] . Through the iterative process of
the ACO algorithms, the agents gradually optimize the task allocation strategy to achieve collaborative coop-
eration in task allocation.

To further optimize the task allocation results, we introduced a GHNN. GHNNs are deep learning models
capable of processing graph-structured data [23] . They can encode the information of nodes and edges in the
graph into vectors and learn the relationships and features between nodes through the neural network learning
process. In heterogeneous multi-agent task allocation, the GHNN takes the state, task features, and graph
structureinformationoftheagentasinputsandoutputstheoptimaltaskallocationstrategyaftercalculationby
the neural network. Through repeated iterative optimization, GHNNs can learn more accurate task allocation
strategies, thereby further improving system performance [24] .

The application of GHNN-ACO algorithms in heterogeneous multi-agent task allocation is a joint optimiza-
tion method [25] . We apply GHNNs and ACO algorithms to heterogeneous multi-agent task allocation, form-
ing a collaborative optimization process [25] . The ACO algorithms are responsible for global search and self-
organization, aiding agents in quickly finding preliminary task allocation schemes. On the other hand, the
GHNNs handle fine learning and optimization of task allocation to obtain more accurate optimal task alloca-
tion strategies [26] . By combining the two approaches, their respective limitations can be overcome, resulting