Page 63 - Read Online

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

in an efficient and accurate task allocation process for heterogeneous multi-agent systems [27] .

Heterogeneous multi-agent systems involve agents with different capabilities, resources, and expertise, neces-
sitating effective task allocation to optimize system performance. While traditional ACO algorithms have
been successfully applied to solve combinatorial optimization problems, they may face challenges in complex
multi-agent environments [28] . To address these issues, we propose a novel paradigm that combines GHNNs
with ACO algorithms to enhance the task allocation process in heterogeneous multi-agent systems [29,30] . The
GHNN-ACO algorithms present an innovative approach to solving the assignment problem in heterogeneous
multi-agent systems [31] . Through the integration of ACO algorithms and GHNNs, this method effectively ad-
dresses heterogeneity and dynamics, achieving efficient and accurate task allocation. We expect this method
to find widespread applications in various fields, such as intelligent transportation systems, robot cooperation,
and distributed sensor networks in the future.

The main contributions are summarized as follows:
1. Use graph neural networks to learn the relationships among heterogeneous multi-agent systems. By con-
structing a graph neural network, tasks and agents are treated as nodes, and edges represent the strength of
the relationship between tasks and agents. The graph neural network can learn the similarity and matching
degree between tasks and agents, providing more meaningful heuristic information for ACO algorithms.
The GHNN-ACO structure can effectively handle heterogeneous graph data, fully utilizing the heterogene-
ity of nodes and edges and modeling different types of nodes and edges separately to improve the adaptabil-
ity of the algorithm to heterogeneous graph data.
2. Graph neural-enhanced ACO algorithms are used to optimize the task allocation process. The graph neural
network enhances the optimal matching between ACO algorithms and search tasks, as ants select the next
task or agent for matching based on the heuristic information learned by the graph neural network. The
pheromone in ACO algorithms represents the degree of execution, adaptability, or collaboration between
tasks and agents. Ants perform task allocation and agent selection according to the pheromone concentra-
tion and the prediction results of the graph neural network to achieve a better task allocation effect.
3. Integrate global and local information and adaptability. The GHNN-ACO algorithm converts graph data
into a low-dimensional vector representation through graph feature learning and fuses global and local
information. Graph feature learning considers the relationship and topology between nodes, enhancing
the representation vector of nodes to contain more semantic information and improving the performance
of classification and prediction tasks. ACO algorithms are adaptive, and path search adjusts path selection
according to pheromone concentration and node characterization information, which helps to find better
paths. The GHNN-ACO algorithm finds better node classification and link prediction solutions in graph
data.
4. Neural networks accelerate the task allocation process of heterogeneous multi-agent systems. The graph-
enhanced neural network is used to accelerate the calculation and decision-making, such as accelerating
the pheromone update in ACO algorithms or the calculation of the graph neural network to improve the
efficiency of task allocation. The GHNN and ACO algorithms of the GHNN-ACO algorithm have a certain
degree of parallelism. The sub-network processes different types of nodes and edges in parallel to speed up
graph feature learning. In ACO algorithms, the ant search paths are independent of each other, and parallel
computing can be used to accelerate the search process.
Through the above innovative combination method, the advantages of graph neural networks and ACO al-
gorithms can be combined to improve the efficiency and accuracy of solving heterogeneous multi-agent as-
signment problems. This method can be applied to complex task allocation scenarios, such as drone group
task allocation, logistics systems task allocation, etc., providing intelligent optimization solutions for practical
application scenarios. In the experiment, it is necessary to fully verify the performance of this method in the
heterogeneous multi-agent assignment problem to ensure that it can effectively solve practical problems.