How to improve energy efficiency (EE) and transmission efficiency for wireless sensor networks (WSNs) through dynamic routing is a challenge. This article introduces MADII, an adaptive routing algorithm integrating the informer model and imitation learning (IL) within a multiagent deep Q-network (DQN) framework. Our method models the routing decision in WSN as a multiagent Markov decision process (MMDP) and employs a novel sensor transmission control mechanism for information via negotiation in centralized-training, centralized-execution and centralized-distribution (STCIN3C) tailored for the centralized training and centralized execution (CTCE) paradigm. This design effectively minimizes redundant energy consumption of sensor nodes (SNs). The reward function of reinforcement learning (RL) is designed by considering transmission quantity, data-receiving success rate, and relative energy consumption. Furthermore, our approach adopts a combination of informer model and IL strategy for balancing exploration and exploitation of RL applications with large-scale action space through deep learning (DL) model and training sample, which can accelerate convergence speed, fit the dynamically changing features of WSN, and obtain better routing scheme. Extensive experiments demonstrate that MADII significantly outperforms existing methods on network lifetime and EE. Furthermore, the ablation experiment emphasizes the effectiveness of informer model and IL strategy in dealing with RL with large-scale action space.
Wireless Sensor Networks (WSNs) have become an indispensable part of the Internet of Things (IoT). How to balance energy consumption and data transmission to further extend the overall lifetime of WSNs remains a key issue for IoT applications deployed on WSN infrastructure. Different from traditional clustering routing algorithms, this paper proposes an adaptive intelligent routing decision algorithm based on deep reinforcement learning (DRL). The proposed model leverages ResNet to extract local spatial features from WSN topology diagram, while BiLSTM captures bidirectional sequence dependencies. These two models are used to address the problems of frequent WSN topology changes and large state spaces. Additionally, the ProbSparse Attention mechanism is adopted to prune large-scale action spaces and make routing decisions. To accelerate the convergence speed of reinforcement learning (RL), the Deep Q-Network (DQN) algorithm incorporates imitation learning (IL) from existing clustering algorithms to balance exploration and exploitation. Experimental evaluations demonstrate the superior performance of the proposed algorithm on network lifespan and energy efficiency. Furthermore, ablation experiments highlight the importance of each model and strategy component.
20 December 2025 in Expert Systems with Applications