Search

Industrial applications introduce new and complex requirements in terms of reliability and latency for wireless communication systems. In particular, 3GPP has recently identified the need for communications being ultra reliable as well as robust against consecutive packet errors. These requirements call for new approaches that span multiple layers to encompass the latency-reliability trade-offs compared to classical error correction schemes like (Hybrid) ARQ. For this purpose, techniques like puncturing, power boosting, and data duplication have been introduced in 5G NR to enable transmission preemption and overriding, and data redundancy. To alleviate their radio inefficiency cost, this paper presents network coding schemes that proactively correct packet errors caused by simultaneous or consecutive leg transmission failures. In particular, we demonstrate that the proposed schemes are able to increase the reliability of single and consecutive packet transmissions while reducing the associated traffic increase as compared to data duplication.

Despite large strides in terms of generative adversarial networks (GANs) for image generation, evaluating and comparing GANs remains an open question. Several measures have been introduced, however, there is no consensus in terms of the best score. In this paper, we delve into the widely-used metric Inception Score (based on KL divergence), revealing that it fails to detect intra-class mode collapse. Meanwhile, Wasserstein distance has received much attention in comparing distributions in recent years but suffers heavy computational burden in high dimensional space. Our idea is that we can find specific embedding space where Euclidean distance could mimic Wasserstein distance to solve the heavy computational problem. This space can be found using a Siamese network, which could be trained quickly because of shared weights. We also apply several proposed new techniques to get better image embedding. To evaluate our proposed metric (Siamese Score), we simulate mode collapse using K-means clustering performed on real data set. To further validate it, we perform an empirical study on several GAN models and use the generated images to do the task. Experiments show that Siamese Score can detect mode collapse and is time-efficient compared with Inception Score and we think our score can be complementary to Inception Score.

Adversarial samples aim to make deep convolutional neural networks predict incorrectly under small perturbations. This paper investigates non-targeted adversarial samples of convolutional neural networks and makes a primitive attempt to characterize adversarial samples. Two observations are made: first, adversarial perturbations are mainly in the high-frequency domain; second, adversarial categories usually have strong semantic relevance to the original categories. Our two observations provide a solid basis to understand the behavior of convolutional neural networks and thus to improve their robustness against adversarial samples.

Generative Adversarial Net is a frontier method of generative models for images, audios and videos. In this paper, we focus on conditional image generation and introduce conditional Feature-Matching Generative Adversarial Net to generate images from category labels. By visualizing state-of-art discriminative conditional generative models, we find these networks do not gain clear semantic concepts. Thus we design the loss function in the light of metric learning to measure semantic distance. The proposed model is evaluated on several well-known datasets. It is shown to be of higher perceptual quality and better diversity then existing generative models.

In this paper, we investigate the use of an integrated machine learning algorithm to jointly optimize the spectrum allocation, load balancing and energy saving aspects in the opportunistic mobile broadband network for temporary event and disaster relief scenarios. A novel k-means algorithm has been developed to dynamically partition the users in a cell into clusters, to improve interference mitigation and spectrum reuse. It is integrated with a Q learning algorithm for resource allocation and transfer learning algorithm for cell selection. Topology management is developed using Q learning to improve BS placement and sleep mode operation. System simulation is carried out using a practical Ljubljana scenario. Compared to the classical LTE resource allocation and cell selection approach, clustered Q learning and transfer learning achieves significant QoS improvement in terms of spectrum and load optimization. With topology management, the learning algorithms show an effective balance between energy saving and QoS.

This paper investigates dynamic topology management functionalities in the flexible deployment of a hybrid aerial-terrestrial cognitive cellular network for public safety in unexpected or temporary events. An evolutionary roll out and roll back of the network architecture in the disaster relief scenario is proposed under the assistant of topology management algorithm, in order to manage the number, location and time of aerial and terrestrial eNBs required by monitoring the user traffic at different phases of the disaster relief operation. Quality of Service (QoS) from user requirements or standards is used to intelligently manage the network topology, which effectively achieves a balance between the deployment cost/energy consumption and QoS/capacity. The results show that with topology management, the dynamic placement strategy based on traffic density significantly improves network QoS compared to fixed placement, and the dynamic deployment strategy based on QoS requirements substantially reduces the scale and power consumption of the network.

This paper presents the concept of the Win-or-Learn-Fast (WoLF) variable learning rate for distributed Q-learning based dynamic spectrum management algorithms. It demonstrates the importance of choosing the learning rate correctly by simulating a large scale stadium temporary event network. The results show that using the WoLF variable learning rate provides a significant improvement in quality of service, in terms of the probabilities of file blocking and interruption, over typical values of fixed learning rates. The results have also demonstrated that it is possible to provide a better and more robust quality of service using distributed Q-learning with a WoLF variable learning rate, than a spectrum sensing based opportunistic spectrum access scheme, but with no spectrum sensing involved.

In this paper, we introduce a novel paradigm of transfer learning for spectrum and topology management in a rapidly deployable opportunistic network for the post disaster and temporary event scenarios. The network architecture is designed to be rapidly changing between different disaster phases, and highly flexible during the temporary event period. Transfer learning is developed to learn the dynamic radio environment from network topologies. This also allows previously learnt information in earlier phases of a deployment to be efficiently used to influence the learning process in later phases of a deployment. A Transfer Learning strategy is designed to change the knowledge base from the most recent phase via multi-agent coordination. We evaluate transfer learning paradigm in a small cell Terrestrial eNB architecture, integrated with Q-Learning and Linear Reinforcement Learning. It is demonstrated that transfer learning significantly improves the initial performance, the convergence speed and the steady state QoS, by exchanging topology information for resource prioritization.

This paper investigates a weighting factor based reinforcement learning scheme with a physical information based channel selection policy applied on a multi-hop backhaul wireless network with directional antennas, for a high capacity density wireless system, in order to enhance the spectrum efficiency and Quality of Service (QoS). The interference environment on a multi-hop backhaul network has been analyzed. A novel channel selection policy is designed based on the interference information obtained from the spectrum sensing process, which is incorporated into a multi-hop based learning scheme. It is demonstrated that the weighting factor based reinforcement learning scheme can efficiently partition channels for users in different locations and achieve a significantly higher QoS than conventional approaches. Moreover, the receiver based interference weighted channel selection policy can speed up the learning process in its initial stage.

supercharged boiler is widely used in large ships as a main power plant because of its small size and weight, high thermal efficiency and volumetric heat load. Based on the fluent software, this paper uses k-e turbulence model and the simplified PDF model to simulate different locations of Steady flame device. Through analyzing the variation tendency of recirculation zone, the flame length, the maximum temperature and the position, the incomplete combustion product of furnace outlet and the average flame temperature with the trend of device location. We can get that the location of Steady flame changes have a significant impact on flame temperature, flame length and the full extent of burning; cold recirculation zone size variation of the combustion conditions and changes of different recirculation zone; when the device's position is in the 139mm, the recirculation zone will have a good size, the flame length will be small and the level of combustion will be well. The results provide a reference for the design and running of supercharger boiler.

In this paper, a novel texture feature GMACM, is presented according to the statistics of gradient angle cooccurrence in color images. Based on three different types of gradients defined in the RGB space, the corresponding GMACMs are introduced. With some well-designed color image classification experiments, it is shown that GMACMs outperform GLCM and Gabor filters significantly in efficiency and accuracy. It could be concluded that GMACM is powerful in classifying and understand color images.

In current interactive segmentation algorithms, im- age models are constructed and simplified to be independent of spatial features of images. This conflicts with receptive field hypothesis of human vision systems, and causes over- segmentation and under-segmentation. Based on receptive field hypothesis, the paper establishes an image modeling method in which spatial distances are taken into account, and a conservative factor is introduced into the image energy function to improve the segmentation veracity. It is shown by experiments that the method is more accurate than its counterparts.

The sensitivity attack is a main threat to the security of watermarking schemes with open detectors. By the attempts across the detection boundary, the attackers gain adequate information of the embedded watermark to remove it with- out introducing serious distortions into the watermarked works. In some image watermarking schemes, the detection boundaries are the patchworks of certain numbers of hyper- planes. Here the existing sensitivity attacks are not suitable anymore for the detection functions have numbers of differ- ent gradients. The letter proposed a new sensitivity attack in which the attacking directions were calculated with a cer- tain number of gradients estimated separately with the old sensitivity attack. Our experiments show the new attack can remove the watermarks successfully without seriously tam- pering the fidelity.