Invalid action masking is a practical technique in deep reinforcement learning to prevent agents from taking invalid actions. Existing approaches rely on action masking during policy training and utilization. This study focuses on developing reinforcement learning algorithms that incorporate action masking during training but can be used without action masking during policy execution. The study begins by conducting a theoretical analysis to elucidate the distinction between naive policy gradient and invalid action policy gradient. Based on this analysis, we demonstrate that the naive policy gradient is a valid gradient and is equivalent to the proposed composite objective algorithm, which optimizes both the masked policy and the original policy in parallel. Moreover, we propose an off-policy algorithm for invalid action masking that employs the masked policy for sampling while optimizing the original policy. To compare the effectiveness of these algorithms, experiments are conducted using a simplified real-time strategy (RTS) game simulator called Gym-$\mu$RTS. Based on empirical findings, we recommend utilizing the off-policy algorithm for addressing most tasks while employing the composite objective algorithm for handling more complex tasks.

Unmanned aerial vehicles (UAVs) have numerous applications in reconnaissance, search, and rescue missions. However, searching underwater areas remains a challenge for UAVs. For anti-submarine search tasks, this paper proposes using a mixed swarm of UAVs and unmanned surface vehicles (USVs). To generate vehicle paths with the goal of minimizing the travel time while achieving complete coverage, a two-phase strategy is developed. In the first phase, optimal waypoints are determined to balance the task volume among the UAVs and USVs. A mapping model is constructed to represent the relationships between the different unmanned platforms and the mission environment. In the second phase, the search trajectory for each UAV/USV is generated by combining waypoint clustering and closed-loop trajectory generation. A detailed simulation study confirms the feasibility and flexibility of the proposed path planning method when applied to cross-domain swarms of different types and sizes for anti-submarine missions. In addition, the simulation results show that cross-domain cooperative planning for unmanned platforms may have advantages over single-domain planning in some scenarios.

This paper investigates a distributed time-varying optimization problem concerning UAV swarm systems with bounded inputs. The objective is to achieve state consensus and minimize the sum of local time-varying objective functions in a distributed manner, while considering the constraints on inputs. To address this problem, we propose a distributed optimization protocol utilizing the projection operator and prediction-correction method. This protocol enables the UAVs to track the time-varying optimal solution with an asymptotically diminishing error. It is shown that the proposed protocol achieves approximate global optimal consensus for UAVs with different connected communication topologies, including undirected, directed detail-balanced, and switching topologies. Consequently, the UAV swarm, equipped with the proposed protocol, can establish a predetermined formation and effectively track moving targets. To validate the theoretical findings, we present three applications: a scalar example, a cooperative hunting task, and swarm tracking behavior.

In the area of artificial intelligence, deep reinforcement learning has grown in significance. It has accomplished extraordinary feats and offers a fresh approach to previously challenging challenges, such as controlling a robotic arm and discovering game strategies. The two primary categories of deep reinforcement learning methods—deep reinforcement learning based on value function and deep reinforcement learning based on policy gradient—are initially explained in this study. The limitations of current approaches and the difficulties faced by deep reinforcement learning methods in related domains are further sorted out, and then the future application directions of deep reinforcement learning methods in the military sphere are examined. Finally, a growing trend for deep reinforcement learning techniques is anticipated in military applications.

This paper proposes a cooperative search algorithm to enable swarms of unmanned aerial vehicles (UAVs) to capture moving targets. It is based on prior information and target probability constrained by inter-UAV distance for safety and communication. First, a rasterized environmental cognitive map is created to characterize the task area. Second, based on Bayesian theory, the posterior probability of a target’s existence is updated using UAV detection information. Third, the predicted probability distribution of the dynamic time-sensitive target is obtained by calculating the target transition probability. Fourth, a customized information interaction mechanism switches the interaction strategy and content according to the communication distance to produce cooperative decision-making in the UAV swarm. Finally, rolling-time domain optimization generates interactive information, so interactive behavior and autonomous decision-making among the swarm members are realized. Simulation results showed that the proposed algorithm can effectively complete a cooperative moving-target search when constrained by communication distance yet still cooperate effectively in unexpected situations such as a fire.