Unmanned aerial vehicles (UAVs) have been widely used in military warfare, and realizing safely autonomous motion control (AMC) in complex unknown environments is a challenge to face. In this paper, we formulate the AMC problem as a Markov decision process (MDP) and propose an advanced deep reinforcement learning (DRL) method that allows UAVs to execute complex tasks in different environments. Aiming to overcome the limitations of the prioritized experience replay (PER), the proposed asynchronous curriculum experience replay (ACER) uses multithreads to asynchronously update the priorities and assigns the true priorities to increase the diversity of experiences. It also applies a temporary pool to enhance learning from new experiences and changes the fashion of experience pool to first-in-useless-out (FIUO) to make better use of old experiences. In addition, combined with curriculum learning (CL), a more reasonable training paradigm is designed for ACER to train UAV agents smoothly. By training in a large-scale dynamic environment constructed based on the parameters of a real UAV, ACER improves the convergence speed by 24.66% and the convergence result by 5.59% compared to the twin delayed deep deterministic policy gradient (TD3) algorithm. The testing experiments carried out in environments with different complexities further demonstrate the strong robustness and generalization ability of the ACER agents.

Learning accurate Bayesian Network (BN) structures of high-dimensional and sparse data is difficult because of high computation complexity. To learn the accurate structure for high-dimensional and sparse data faster, this paper adopts a divide and conquer strategy and proposes a block learning algorithm with a mutual information based K-means algorithm (BLMKM algorithm). This method utilizes an improved K-means algorithm to block the nodes in BN and a maximum minimum parents and children (MMPC) algorithm to obtain the whole skeleton of BN and find possible graph structures based on separated blocks. Then, a pruned dynamic programming algorithm is performed sequentially for all possible graph structures to get possible BNs and find the best BN by scoring function. Experiments show that for high-dimensional and sparse data, the BLMKM algorithm can achieve the same accuracy in a reasonable time compared with non-blocking classical learning algorithms. Compared to the existing block learning algorithms, the BLMKM algorithm has a time advantage on the basis of ensuring accuracy. The analysis of the real radar effect mechanism dataset proves that BLMKM algorithm can quickly establish a global and accurate causality model to find the cause of interference, predict the detecting result, and guide the parameters optimization. BLMKM algorithm is efficient for BN learning and has practical application value.

As an increasing number of intelligent decision-making problems of vehicles are addressed using implementations of deep learning (DL) methods, the interpretability of intelligent decision-making strongly determines the degree of human trust in the technology and implementation process; therefore, it is urgently necessary to improve the interpretability of DL. To enhance the interpretability of intelligent decision-making processes based on DL, we propose a feature explanation network (FENet), which consists of a correlation clustering module, a feature extraction module, and a decision module. First, to determine the source of each feature before the feature extraction module, input variables are clustered according to correlation. We design dynamic strategy and Dset strategy, the former strategy comprises four specific implementations corresponding to different parameter selections and noise processing methods. Then, in the feature extraction module, any DL model can be employed to extract features from each correlation cluster separately. Finally, a probabilistic graphical model-Bayesian network is constructed based on the features, and the intelligent decision is realized by Bayesian inference. In the experiments, we first design five indicators to quantify the interpretability and verify the effectiveness of each module in FENet by an ablation study. Compared with other DL models, FENet demonstrates significant improvements in interpretability while maintaining comparable decision accuracy. Moreover, when applied to the practical problem of UAV intrusion detection, FENet can provide an explainable process and interpretation of final results. These findings indicate that the proposed network architecture can be effectively utilized with various DL-based algorithms for decision-making problems to enhance interpretability.

Restricted Boltzmann machines (RBMs) are commonly used as pre-training methods for deep learning models. Contrastive divergence (CD) and parallel tempering (PT) are traditional training algorithms of RBMs. However, these two algorithms have shortcomings in processing high-dimensional and complex data. In particular, the number of temperature chains in PT has a significant impact on the training effect, and the PT algorithm cannot fully utilize parallel sampling from multiple temperature chains for the divergence of the algorithm. The training can quickly converge with fewer temperature chains, but this impacts the accuracy. More temperature chains can help PT achieve higher accuracy in theory, but severe divergence at the beginning of the training may ruin the training result. To exploit fully the advantages of PT and improve the ability of RBMs to process high-dimensional and complex models, this paper proposes dynamic tempering chains (DTC). By dynamically changing the number of temperature chains during the training process, DTC starts training with fewer temperature chains and gradually increase the number of temperature chains with training going on, and finally get an accurate RBM. And one-step reconstruction error is proposed to measure the convergence, which can decrease the influence of the dynamic training strategy on reconstruction error. Experiments on MNIST, MNORB, Cifar 10, and Cifar 100 indicate that, compared with PT, the classification accuracy of DTC algorithm improved by up to 8%. DTC quickly converges in the early stage of training because of few exchanges among temperature chains and produces higher accuracy at the end for the global optimum model learned by more temperature chains, especially when learning high-dimensional and complex data. This proves that the DTC algorithm effectively utilizes parallel sampling of multiple temperature chains, overcomes divergence challenges, and further improves the training effect of the RBM.