Conventional routing algorithms are typically divided into static and dynamic categories. Static routing involves manually configured routing tables and is incapable of adapting to variations in network topology or traffic conditions. As networks grow in scale and complexity, manual configuration becomes increasingly cumbersome and error-prone, rendering static routing unsuitable for dynamic environments. In contrast, dynamic routing algorithms autonomously adjust routing tables in response to network conditions, making them more adaptable and suitable for complex, dynamic communication scenarios (Zhou et al., 2023).

Traditional routing algorithms are supported by well-established theoretical foundations and have been widely applied across various domains. For instance, Mehraban et al. introduced an intelligent node placement strategy that integrates both node connectivity and traffic features, demonstrating notable improvements in reducing maximum link utilization over existing methods (Hussain and Sinha Roy, 2021). Similarly, Zhu et al. introduced a routing algorithm tailored for low Earth orbit satellite ad hoc networks by enhancing open shortest path first (OSPF) to accommodate dynamic satellite topologies, incorporating a routing reconstruction mechanism to ensure reliable data transmission (Zhu and Jin, 2023). Hussain et al. introduced a smart flow steering agent that leverages real-time network conditions to dynamically balance traffic in SDN, thereby enhancing routing efficiency and overall performance (Hussain et al., 2021). However, as network traffic and scale continue to grow exponentially, conventional routing algorithms face mounting limitations due to their restricted ability to obtain real-time link state information. This limitation results in slower convergence and delayed responsiveness to dynamic traffic changes, thereby reducing transmission efficiency. Consequently, Traditional routing strategies are inadequate for addressing the complex and dynamic demands of tactical communication networks in intelligent combat scenarios, thereby hindering the formulation of globally optimal routing decisions.

Heuristic algorithms are well-suited for solving complex optimization problems and offer notable advantages over traditional methods when addressing large-scale and dynamic network environments. As such, they have increasingly become mainstream solutions for routing optimization tasks. Tabatabaei et al. combined the bacterial foraging optimization algorithm with a mobile sink approach to propose a novel routing algorithm for wireless sensor networks. By incorporating node energy information in the path selection process, the algorithm effectively reduced control overhead and bandwidth consumption while improving routing accuracy (Tabatabaei, 2022). Alnajjar et al. developed a genetic algorithm to optimize cluster head placement in wireless sensor networks, achieving balanced distribution and stable energy usage. By employing customized crossover and mutation operations, the method gradually improved sensor-to-cluster assignments, confirming the efficiency and reliability of the approach over multiple generations (Alnajjar et al., 2022). Deng et al. (2022) introduced an ant colony optimization (ACO)-based routing algorithm for low Earth orbit (LEO) satellite networks. Their method jointly considers path distance, transmission direction, and link load to identify routes with minimal delay and cost, thereby achieving load balancing in the satellite network. Shokouhifar (2021) designed a fuzzy heuristic ACO algorithm that utilizes a multi-objective function incorporating power consumption, delay, and reliability, allowing for adaptive routing based on application-specific requirements. Ju et al. designed a dynamic load-balancing routing approach for LEO satellite networks was developed by integrating multiagent dueling double deep Q-network with spatiotemporal traffic prediction. By utilizing both local and predicted traffic data, the method efficiently optimizes routing, reduces peak and average bandwidth usage, and balances network load (Ju et al., 2025). While heuristic routing algorithms have demonstrated strong capability in solving complex network optimization problems and have been widely applied, they often require specific network assumptions and conditions. Variations in network topology and link status can introduce significant volatility and error, undermining algorithm scalability and stability. As a result, these methods tend to perform poorly in highly dynamic and unpredictable environments, such as tactical communication networks under battlefield conditions, where rapid changes in structure and link quality are common. Consequently, heuristic routing algorithms face challenges in maintaining robust QoS under such circumstances.

DRL offers the capability to learn complex, nonlinear policies through interaction with the environment, exhibiting superior flexibility and adaptability in solving practical problems. These advantages have opened new avenues for addressing complex network optimization challenges (Dake et al., 2021). To address the limitations of traditional routing strategies in SDN, particularly their inability to meet diverse application-level QoS requirements and their lack of interaction with the network environment, Li et al. (2021) proposed DRNet—an intelligent, QoS-aware routing algorithm. This approach utilizes traffic demand matrices and network state information as input to the DRL agent for training, enabling more flexible routing deployment. Fu et al. (2020) designed a deep Q network (DQN)-based routing strategy that distinguishes between mouse flows and elephant flows in data center networks, thereby improving throughput and reducing latency and packet loss. Liu et al. (2021) introduced a DRL-based routing algorithm that reconfigures network resources by assigning weighted importance to caching, bandwidth, and latency. Network states are represented as multidimensional inputs to the agent, enabling perceptual learning and optimized resource allocation, which enhances throughput and robustness. Casas-Velasco et al. (2021) proposed a DQN-based intelligent routing algorithm under the SDN architecture that learns path state information for dynamic routing decisions, addressing the challenge of traffic variability in traditional routing schemes. He et al. integrates graph neural networks with deep reinforcement learning to capture network topology information and guide routing decisions, which enables balanced traffic distribution and enhanced performance (He et al., 2023). Huang et al. (2022) developed a deep graph reinforcement learning (DGRL) routing algorithm that integrates graph convolutional network (GCN) with DDPG to achieve effective traffic control in wireless sensor networks. Zhou et al. (2022) proposed a PPO-based QoS-aware routing optimization mechanism, which adjusts the reward function dynamically according to different optimization goals, thus enabling a generic and customizable routing optimization framework.

These DRL-based intelligent routing algorithms demonstrate strong adaptability to complex and dynamic network conditions, enabling autonomous policy adjustments and improving overall service quality. However, several limitations persist. Current methods often suffer from slow convergence and unstable performance, and many focus solely on single or limited link-level features, neglecting the heterogeneous QoS demands across different traffic types. Moreover, the highly dynamic topology and rapidly changing link states of tactical communication networks challenge the fixed-architecture neural networks used in existing models, limiting their situational awareness and adaptability. Consequently, current intelligent routing algorithms are inadequate for optimal path decision-making in highly mobile, adversarial battlefield environments.

This study employs Ryu (Ryu SDN Framework Community, 2020) as the SDN controller to achieve network topology awareness and link status monitoring, implementing the described application requirements using Python programming.

The PPO algorithm is an improvement on the trust region policy optimization (TRPO) algorithm (Li and He, 2023). Its main advantage is to limit the distance between the new policy π θ and the old policy π θ o l d in each update to ensure that the policy update amplitude is not too large, thereby improving the stability of the algorithm.

The basic framework of the PPO algorithm can be divided into two parts, the Actor network and the Critic network. The Actor network is a policy network that maps the state s t at time t to the probability distribution of the action a t . While the Critic is a value network that evaluates the value of the current state s t . In the traditional policy gradient update algorithm, after the agent collects experience samples according to the strategy π θ , the parameter θ will be updated. Then the agent will re-collect experience samples according to the new strategy, which will lead to a decrease in sample utilization. To solve this problem, the PPO algorithm uses the importance sampling method shown in Equation 5 to limit the policy update range, thereby improving the utilization of experience samples, where π θ a t s t and π θ o l d a t s t represent the new and old policy probabilities of taking action s t under state a t . r t θ = π θ a t s t π θ o l d a t s t (5)

The PPO algorithm can be refined into PPO-penalty with adaptive KL-divergence and PPO-clip with clip clipping according to the policy update method of the Actor network. The objective functions of PPO-Penalty and PPO-Clip are defined in Equations 6, 7. L k l p e n θ = E t r t θ A ^ t − β K L π θ o l d · s t , π θ · s t (6) L c l i p θ = E t min r t θ A ^ t , c l i p r t θ , 1 − ε , 1 + ε A ^ t (7) where A ^ t represents the advantage function, which is used to measure the quality of the agent’s actions in the policy network. β is the penalty term coefficient. ε represents the clipping factor. Studies have shown that PPO-Clip has more efficient sample utilization and feasibility than PPO-Penalty.

The intelligent optimization algorithm proposed in this paper uses PPO to construct a deep reinforcement learning agent for model training. PPO is more versatile and stable than DDPG, actor-critic and other algorithms (Li and Wu, 2025). The PPO algorithm is implemented based on the AC network architecture. The policy network Actor is used to learn the mean of the network environment state output action, and the evaluation network Critic is used to output the state value. The agent interacts with the environment through state, action and reward information, and continuously optimizes the strategy through iterative training.

As a framework-based deep reinforcement learning algorithm, PPO requires the design of different state spaces, action spaces and reward functions for different problems and different application scenarios. In this paper, based on the PPO deep reinforcement learning framework, the state space, action space and reward function are designed for the tactical communication network scenario, as follows:

1. State space. The state space is the set of link state information obtained by the agent from the tactical communication network. The network state x t at any time t includes the global network service request Q E t , the link remaining bandwidth B E t , the delay D E t and the packet loss rate L E t information, as shown in Equation 11.

The specific representation of each network link state information set is shown in Equations 13–15. Due to the large differences in the element values in the link state information, this difference will cause the intelligent path optimization algorithm to fluctuate greatly during the training process and it is difficult to reach a convergence state. Therefore, the Min-Max method is used to normalize the elements in the link state information. B E t = b e 12 t , b e 23 t , ⋯ , b e i j t , i , j ∈ m (13) D E t = d e 12 t , d e 23 t , ⋯ , d e i j t , i , j ∈ m (14) L E t = l e 12 t , l e 23 t , ⋯ , l e i j t , i , j ∈ m (15)

The state space of the agent is designed to be a state sequence formed by combining the state information containing time t and l−1 time steps before the time, as shown in Equation 16. s t = x t − l + 1 , x t − l + 2 , ⋯ , x t ∈ R l × d × n (16)

2. Action space. The action space refers to the set of actions taken by the agent according to the state s t during the interaction with the network environment, which is used to forward the business flow in the network. Assume that each source-destination node pair in the global network contains k feasible paths to form a feasible path set P = p 1 , p 2 , p 3 , ⋯ , p k , where each path corresponds to a weight coefficient, namely, W P = w 1 p , w 2 p , ⋯ , w k p . The algorithm in this paper defines each action in the action space as the forwarding path weight between all source-destination node pairs in the global network, specifically expressed as a t = w 12 , ⋯ , w i j , ⋯ , w s d . Among them, w i j ∈ W P , i ∈ 1 , s , j ∈ 1 , d represent the weight of the forwarding path p selected between the source node i and the destination node j, and w i j ∈ 0 , 1 .

The PPO algorithm policy update stage uses the advantage function A ^ t to measure the quality of each action. The advantage function is defined as follows: A ^ t = R t + γ · V s t + 1 − V s t (18) where R t is the reward for executing action a t in state s t , γ represents the reward discount factor, V s t and V s t + 1 represent the evaluation values of the current state s t and the next state s t + 1 respectively.

The traditional policy gradient update algorithm will cause serious deviations in the execution strategy of the agent due to the inaccurate estimation of the advantage function [33]. Therefore, this paper adopts the importance sampling method to adjust the strategy update amplitude and improve the efficiency of strategy update and sample utilization during training. The importance sampling is presented in Equation 19. r t θ = π θ a t s t π θ o l d a t s t (19) where r t θ represents the ratio of the probability that the current strategy π θ takes action a t in state s t to the probability that the old strategy takes action a t in state s t .

To better adapt to the dynamic changes of the network topology and the uneven distribution of samples, the algorithm in this paper adopts a gradient clipping method defined as Equation 20, which effectively limits the magnitude of policy updates. The policy update of the algorithm is shown in Equation 21. While restricting the extent of policy updates, the algorithm also aims to maximize the expected cumulative reward, thereby improving the convergence and stability of the algorithm. L c l i p θ = E t min r t θ A ^ t , c l i p r t θ , 1 − ε , 1 + ε A ^ t (20) θ ^ = argmax E t min r t θ A ^ t , c l i p r t θ , 1 − ε , 1 + ε A ^ t (21) where, θ ^ represents the updated policy parameters. ε denotes the clipping factor, usually represented by a small positive number, which limits the extent of policy updates within the range of 1 − ε , 1 + ε .

In order to better extract the temporal and spatial characteristics of the tactical communication network link state information, this paper introduces GRU, GAT, and a self-attention mechanism to construct a link state feature extraction module, replacing the feedforward neural network structure of PPO. This enables the perception of the spatiotemporal characteristics of link state information and realizes end-to-end optimal action policy output. As shown in Figure 6, the link state feature extraction module mainly consists of a temporal feature extraction module, a self-attention mechanism module, a spatial feature extraction module, and a multilayer perceptron module.

The experimental scenario constructs an integrated tactical communication network oriented toward joint operations. The network structure is shown in Figure 8, consisting of 47 nodes and 61 links. Among them, nodes 18, 19, and 20 and nodes 32, 33, and 34 simulate sensor nodes. Nodes 42, 43, and 44 simulate command and control nodes. Nodes 11, 12, and 13, nodes 25, 26, and 27, and nodes 39, 40, and 41 simulate fire strike nodes. The traffic transmission path follows the principle of “Sensor — Command and Control — Fire Strike.” The tactical communication network contains more than 10 types of heterogeneous links. The types of heterogeneous links are shown in Table 2. To simulate the impact of a highly adversarial environment on link status, experiments are conducted by setting different link bandwidths in the Mininet simulation software.

The experiment sets five different traffic intensities in units of kbps: {20, 40, 60, 80, 100} for testing. For each traffic intensity, an Iperf [34] script is written to implement one-to-one or one-to-many traffic transmission from sensor nodes to fire strike nodes, and the Ryu controller is used to measure network performance parameters.

The parameter settings of the DRL-SGA algorithm are shown in Table 3.

To verify the effectiveness of the DRL-SGA algorithm in improving the quality of service under different traffic intensities, this paper compares the DRL-SGA algorithm with the following three routing algorithms:

1. OSPF: Open shortest path first algorithm. It obtains the weight information of each link in the network through an SDN measurement mechanism and calculates the path with the shortest link weight.

The experiment uses average network throughput, average end-to-end delay, and average packet loss rate as evaluation metrics. For each traffic intensity, data transmission is tested five times, and the average of the measurement results is taken for comparison. The results are as follows:

The comparison results in Figure 9 indicate that under low traffic intensity scenarios, the throughput improvement of DRL-SGA over other algorithms is not significant. This is because the low service traffic intensity and sufficient link bandwidth resources do not lead to noticeable network congestion, resulting in relatively small differences in throughput across algorithms. However, as traffic intensity increases, DRL-SGA demonstrates a more pronounced advantage in throughput performance compared to the other algorithms. In high traffic intensity scenarios, DRL-SGA consistently maintains a relatively high level of network throughput. When the traffic intensity reaches 100 kbps, the throughput improves by up to 23.48% compared with the OSPF algorithm, and by up to 2.59% compared with the better-performing DRL-ST algorithm. This indicates that DRL-SGA can formulate more optimal routing strategies based on network load conditions and dynamic link state variations.

As shown in Figure 10, DRL-SGA achieves lower average end-to-end delay than the comparison routing algorithms under all traffic intensity levels. Moreover, as the traffic intensity increases, DRL-SGA demonstrates more outstanding performance in guaranteeing end-to-end latency. Compared with the traditional OSPF algorithm, DRL-SGA reduces the average end-to-end delay by at least 14.42%, with a maximum reduction of 33.57%. Compared with the DQN algorithm, the delay is reduced by a minimum of 5.69% and up to 25.04%. When compared with the DDPG algorithm, the reduction ranges from 7.08% to 22.44%. Even compared to the better-performing DRL-ST algorithm, DRL-SGA achieves a delay reduction of at least 2.07% and up to 16.88%.

The comparison results in Figure 11 show that DRL-SGA consistently achieves a lower average packet loss rate than the comparison routing algorithms across different traffic intensities. Moreover, as traffic intensity increases, DRL-SGA maintains more stable packet loss performance. Compared with the traditional OSPF algorithm, the average packet loss rate of DRL-SGA is reduced by at least 14.66%. When compared with intelligent algorithms DQN, DDPG, and DRL-ST, the average packet loss rate is reduced by at least 9.03%, 8.73%, and 1.65%, respectively.

The above results demonstrate that the DRL-SGA algorithm exhibits superior adaptability to network environments with varying traffic intensities. This advantage stems from the fact that OSPF is a routing algorithm based on fixed forwarding rules, which makes it difficult to dynamically adjust routing strategies in response to changes in network traffic, thereby increasing the likelihood of link congestion. In contrast, DQN, DDPG, and DRL-ST utilize deep reinforcement learning mechanisms to perceive network state changes in real time and dynamically adjust optimal routing strategies. This capability allows them to alleviate network congestion to a certain extent, thereby reducing latency and packet loss while improving throughput compared to the OSPF algorithm. However, due to the limitations of fully connected feedforward neural networks used in these traditional methods, their perception capabilities are constrained, which hinders optimal path decision-making.

DRL-SGA, by leveraging the collaborative functionality of GRU, GAT, and attention mechanisms, effectively captures the spatiotemporal characteristics of network link states. It fully exploits hidden information within the link state data and captures evolving trends in link status, thereby enhancing the model’s decision-making and perception capabilities and enabling the generation of more optimal routing strategies. As a result, DRL-SGA demonstrates robust adaptability to varying traffic intensity conditions, effectively reducing end-to-end delay and packet loss while improving network throughput, thereby ensuring real-time and reliable transmission of service data.

To evaluate the effectiveness of individual components in the link-state feature extraction module for enhancing the perceptual capability of the DRL-SGA algorithm, an ablation study was conducted under the scenario where three links were broken and traffic intensity was set to 100 kbps. The attention mechanism, the temporal feature extraction module (GRU), and the spatial feature extraction module (GAT) were removed individually. The performance of each ablated algorithm was compared with that of the full DRL-SGA model by analyzing the convergence behavior of their reward functions. The comparative results are illustrated in Figure 12.

As shown in Figure 12, the removal of any component from the link-state feature extraction module resulted in a decrease in the final reward value at convergence, indicating a reduction in the perceptual and decision-making capabilities of the agent model. This led to suboptimal routing choices. Furthermore, the removal of each module increased the number of iterations required for convergence and slowed the convergence speed. The most significant degradation was observed when the GAT module was ablated, which yielded the lowest reward and slowest convergence. This is attributed to the role of GAT in capturing spatial dependencies in link states, which enhances the algorithm’s adaptability to dynamic network changes and is crucial for achieving effective convergence. Table 4 presents a quantitative comparison of the mean and variance of reward values obtained during training under each ablated condition.

The results indicate that, while the variance in rewards remains comparable across configurations, the full DRL-SGA algorithm achieved mean rewards that are 2.60%, 1.75%, and 1.11% higher than those of the models without GAT, GRU, and the attention mechanism, respectively. This demonstrates that each component in the link-state feature extraction module contributes positively to enhancing the perceptual ability of the agent, enabling the algorithm to adapt to dynamic structural changes in the network and thereby facilitating the output of an optimal routing policy.

To further assess the feasibility of the proposed DRL-SGA algorithm in practical applications, its computational overhead and scalability were analyzed. Evaluation metrics included model size, parameter count, per-epoch training time, inference latency, peak memory usage, throughput, and control overhead. The experimental environment was identical to that described earlier, and synthetic networks of different scales were constructed to simulate various deployment scenarios: small scale (16 nodes, 24 links), medium scale (47 nodes, 61 links), and large scale (100 nodes, 220 links). The results are presented in Table 5.

Several conclusions can be drawn from Table 5. First, both model size and parameter count remain largely stable across different network scales, indicating that computational complexity is primarily determined by the architecture components (GRU, GAT, and MLP) rather than node count. Second, per-epoch training time grows approximately linearly with network size, consistent with the complexity of node-based feature extraction. Third, even in the large-scale scenario with 100 nodes and 220 links, inference latency remains below 10 ms, meeting the real-time requirements of tactical communication environments. Peak memory consumption increases with network size but remains within the capacity of a V100 GPU. Throughput decreases as the network expands, yet continues to support several hundred routing decisions per second, sufficient for operational needs. Meanwhile, control-plane overhead rises nearly linearly with network size but remains negligible compared with conventional network telemetry traffic. In summary, the DRL-SGA algorithm demonstrates moderate computational overhead and strong scalability. Its complexity scales nearly linearly with network size, making it suitable for deployment in large-scale tactical communication networks. Furthermore, additional strategies such as batch processing, localized attention, and model compression (e.g., pruning or quantization) may further reduce computational costs in resource-constrained environments. Overall, DRL-SGA achieves a favorable balance between routing performance and computational efficiency, supporting its feasibility for real-world large-scale deployment.

As shown in Table 2, the performance of DRL-SGA is influenced by several parameters, among which ε, α, β, and γ are the most critical. To validate the rationality of their selection, two sets of experiments were conducted. In the first, β and γ were fixed while ε and α were varied. In the second, ε and α were fixed while β and γ were adjusted. The results are illustrated in Figure 13.

In the first set, ε was varied within {0.10, 0.15, 0.20, 0.25, 0.30}, and α within {0.05, 0.10, 0.15, 0.20, 0.25}. The best performance was observed when ε = 0.20 and α = 0.10, achieving a maximum average reward of 87.29. This indicates that a balanced exploration rate and reward weighting effectively enhance routing decisions. Excessively high or low exploration rates reduced performance, suggesting that both over-exploration and premature convergence degrade effectiveness. In the second set, β and γ were varied within {0.05, 0.10, 0.15, 0.20, 0.25}. Optimal performance was obtained at β = 0.10 and γ = 0.10 (maximum reward 87.29), with only marginal performance differences across values. This demonstrates robustness in weighting for delay and packet-loss optimization. However, further increases in β or γ caused slight declines, indicating that overly emphasizing a single objective weakens overall multi-objective optimization. Overall, DRL-SGA exhibits a “middle-value optimal” pattern across ε, α, β, and γ, with stable and peak performance maintained under balanced parameter settings. The chosen configuration is therefore reasonable, providing an effective trade-off among exploration, delay optimization, and packet-loss control, ensuring both effectiveness and robustness across diverse network scenarios.

To evaluate the adaptability of the DRL-SGA algorithm to dynamically changing network topologies, a new network structure was configured as shown in Figure 14. Several links within the backbone network were sequentially disconnected to simulate link failure scenarios. Experiments were conducted under both low-intensity traffic (50 kbps) and high-intensity traffic (100 kbps), and the network performance metrics of each routing algorithm were recorded and compared accordingly.

Figure 15 presents the throughput comparison results of each algorithm under varying levels of link failures and different traffic intensities. As the number of failed backbone links increases, the overall network throughput of all algorithms gradually declines. However, the DRL-SGA consistently maintains the highest throughput performance. When the number of failed backbone links reaches three, under the low traffic intensity scenario (50 kbps), the average throughput of DRL-SGA improves by 8.76% over the traditional routing algorithm OSPF, and by 5.34%, 4.25%, and 2.77% over the intelligent routing algorithms DQN, DDPG, and DRL-ST, respectively. Under the high traffic intensity scenario (100 kbps), DRL-SGA outperforms these four algorithms with throughput improvements of 35.29%, 25.47%, 20.18%, and 4.39%, respectively.

Figure 16 shows the comparison of average end-to-end delay across algorithms under different levels of link failure and traffic intensities. As the number of failed links increases, all algorithms experience an increase in delay. Under 50 kbps traffic intensity, the average end-to-end delay of DRL-SGA is at least 12.99%, 8.06%, 6.41%, and 3.15% lower than that of OSPF, DQN, DDPG, and DRL-ST, respectively. Under 100 kbps traffic intensity, DRL-SGA achieves delay reductions of at least 24.25%, 16.95%, 14.43%, and 6.44% compared to the same respective algorithms.

Figure 17 illustrates the comparison of average packet loss rates for all algorithms under increasing link failure conditions and varying traffic intensities. As the number of failed links grows, all algorithms exhibit rising packet loss rates. When the traffic intensity is 50 kbps, the average packet loss rate of DRL-SGA is reduced by at least 21.92%, 15.08%, 12.52%, and 1.65% relative to OSPF, DQN, DDPG, and DRL-ST, respectively. Under 100 kbps traffic intensity, DRL-SGA achieves minimum reductions of 35.19%, 26.44%, 21.16%, and 9.63% compared to the same algorithms.

In summary, compared with traditional routing algorithms and existing intelligent routing algorithms, DRL-SGA demonstrates superior adaptability to dynamic network changes and incurs smaller QoS performance degradation in scenarios involving link failures, regardless of whether the traffic intensity is low or high. This is primarily because the traditional OSPF algorithm is incapable of promptly adapting to changes in network topology and adjusting routing strategies in a timely manner, making it difficult to maintain optimal network performance. Although intelligent algorithms such as DQN, DDPG, and DRL-ST leverage deep reinforcement learning, they rely on conventional feedforward neural networks trained with link state information in fixed formats, making it difficult to adapt to complex network dynamics. Consequently, these algorithms exhibit lower training efficiency and suboptimal decision accuracy.

In contrast, the DRL-SGA algorithm effectively extracts the spatiotemporal characteristics of tactical communication network structures. Specifically, GRU captures the temporal evolution of historical link state information, enabling accurate prediction of future trends and enhancing the model’s perception and reasoning capabilities. GAT enhances the algorithm’s ability to represent non-Euclidean graph-structured data, thereby enabling a more comprehensive understanding and processing of complex network environments and significantly improving the model’s structural awareness. As a result, DRL-SGA exhibits stronger adaptability to dynamic network topologies and superior capability in ensuring service quality.