Speculative decoding has emerged as a promising paradigm for accelerating large language model inference by leveraging a lightweight draft model to generate multiple candidate tokens.However, existing methods often incur substantial training overhead to mitigate information misalignment between autoregressive draft model training and decoding.To address this challenge, we propose EDSD, an Entropy-Driven Speculative Decoding framework that uses entropy as a unified, interpretable signal for both draft model training and architectural design.EDSD drives the draft model to progressively align with the target model in an easy-to-hard manner while establishing tokenlevel alignment as a dominant design principle.Extensive experiments on seven LLMs demonstrate that EDSD improves training efficiency by 24.8%, increases the average acceptance length by 4.0%, and achieves a 4.1% speedup compared to state-of-the-art methods.Furthermore, EDSD improves robustness to system prompt variations by more than 5×.Our findings establish entropy-driven alignment as an effective and principled foundation for efficient speculative decoding.We make our draft model weights available at https://github.com/KerwinKai/EDSD.

This paper presents a power-efficient discrete-time (DT) delta-sigma analog-to-digital converters (ΔΣ ADC) using a correlated level shifting (CLS)-assisted fast self-quenching floating inverter amplifier (FIA) with a sampling noise cancellation (SNC) technique. The CLS technique improves the equivalent open-loop gain of the FIA, while the SNC technique minimizes the sampling noise of the 1st-stage integrator. Compared with the conventional FIAs, the fast self-quenching FIA reduces power consumption while ensuring adequate gain. The prototype of the 2nd-order DT ΔΣ ADC with the proposed FIA is fabricated in a 65-nm CMOS process and achieves fully dynamic operation. It realizes an 88.0dB peak signal-to-noise-and-distortion ratio (SNDR) with an oversampling ratio of 256 for a 19.5kHz bandwidth while consuming 76.2µW from a 1.2V supply at a 10MHz sampling frequency. This ΔΣ ADC achieves 89.6dB dynamic range and 172.1dB SNDR-based Schreier figures of merit.

For the problem of multi-UAV collaborative anti-jamming communication in broadband and complex dynamic jamming environments, this paper proposes a method based on deep reinforcement learning to optimize the communication channel selection of UAV clusters. In order to address the difficulty in learning collaborative anti-jamming strategies caused by multiple jamming patterns in broadband communication, we construct a hierarchical collaborative learning mechanism. The cluster head UAV employs a band selection neural network to choose sub-bands, while the cluster member UAVs perform channel access within the selected sub-band using a channel selection neural network. Each cluster adopts distributed learning, eliminating the need for information exchange between clusters. By decomposing the large action space into two subproblems of band selection and channel access, the learning process is accelerated and the convergence speed is improved. Simulation results demonstrate that this algorithm can rapidly learn effective communication strategies, enhancing the system’s anti-jamming capability and improving overall stability.

This paper investigates a novel anti-eavesdropping strategy based on unmanned aerial vehicle (UAV)-mounted simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS). In particular, a UAV equipped with a STAR-RIS acts as a passive relay to reflect desired signals and simultaneously acts as a friendly jammer to transmit artificial noise (AN) against eavesdroppers. Based on the phase shift coupling characteristics of STAR-RIS, three phase shift configuration strategies are proposed, namely reliability-priority (RP), security-priority (SP), and element-partitioning (EP) schemes. Analytical closed-form expressions of connection outage probability (COP), secrecy outage probability (SOP), effective secrecy throughput (EST) and secrecy energy efficiency (SEE) are derived to evaluate the reliable and secure performance achieved by the proposed schemes, respectively. The asymptotic analysis is also performed for further insights. Analysis and simulation results demonstrate that the proposed three schemes outperform traditional benchmark schemes. From the perspective of reliability, the RP scheme can achieve the best COP. In terms of security, as the number of STAR-RIS elements increases, the SOPs of the SP and EP exponentially decrease, whereas the SOP of the RP scheme increases. The EP scheme achieves the optimal EST, and the asymptotic EST is independent of phase estimation errors. Additionally, it is recommended that the UAV be deployed near the eavesdropper for the SP and EP schemes to enhance SEE.

With the rapid development of reinforcement learning (RL)-enhanced anti-jamming wireless communication technologies and jamming technologies, intelligent communication confrontation has become an urgent problem to be solved. Most existing work assumed that detailed information of jammer was known in advance, which hardly holds in practice. Besides, some work was sensitive to the changing of jamming policy, leading to limited adaptability and scalability. This article extends the research to scenarios with unknown jammer and diverse jamming policies, including fixed, reactive, and deep RL (DRL)-based proactive jamming policies. The interaction between communication party and jammer is formulated as a partially observable adversarial team stochastic game (POATSG). To cope with unknown and diverse jamming policies, a prioritized fictitious rival play (PFRP)-based robust anti-jamming spectrum access scheme (RASAS) is proposed. First, a fictitious jammer is designed to force the communication party to promote robustness via adversarial training. Then, a synchronized update mechanism is adopted to mitigate the nonstationary issue. Finally, the fictitious agent pool is introduced to create diverse fictitious opponents and avoid overfitting. Simulation results show that the PFRP-based scheme is robust to the jamming policy, switching cycle of the jamming policy, and jamming channel number.

This paper focuses on distributed resource management and task scheduling for multi-access MEC networks against the intelligent eavesdropping jammer (IEJ). Due to the lack of a central controller, the problem of joint task scheduling and network resource allocation is formulated as a distributed multi-user hybrid-integer non-convex model.The optimization objective is to maximize users’ satisfaction while meeting the Quality of Service (QoS) requirements of tasks and ensuring the high-reliable demands of data offloading. To overcome the challenge of partial observability for users, the channel observation matrix and Gramian Angular Field (GAF) are utilized to preprocess the limited channel state information and to mine the potential time-frequency characteristics of the external environment. Moreover, the hierarchical architecture and parallel networks are introduced for a parameterized redesign of the Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) method to improve the decision accuracy. Finally, simulation results demonstrate the superiority of the proposed algorithm over existing methods in terms of delay, energy consumption, and security.

This paper studies the anti-jamming channel selection problem in the unmanned aerial vehicle (UAV) communication scenario using machine learning. Recently, deep reinforcement learning (DRL) based anti-jamming approaches have drawn much attention, but most of them require lots of computing resources and power supply for training, which is impractical for the hardware-limited UAVs. What's more, the high complexity of DRL-based algorithms weakens their online learning ability, failing to rapidly adapt to the changing jamming environment. To be applicable to the hardware-limited UAVs, we propose a lightweight reinforcement learning algorithm based on the idea of spectrum state abstraction. We first assign similar spectrum states to clusters using the DRL and clustering algorithms. A state clustering network is deployed in the UAV to convert the large and redundant state space into a small number of state clusters. Based on the clustered states, the UAV uses a simple tabular Q-learning algorithm to online find the optimal anti-jamming policy. The simulation results show that, compared with the conventional DRL approach, the proposed algorithm can efficiently find the optimal anti-jamming policy and fast adapt to the change of jamming pattern in the complicated and dynamic jamming environment.

Sensor deployment and spectrum data completion are two critical stages in radio map construction, which are intricately linked and jointly influence both efficiency and performance. However, existing research often addresses these two issues separately, neglecting their joint optimization. In this study, we introduce two heuristic algorithms to tackle the joint optimization problem of these issues. These algorithms utilize a novel approach that combines deterministic search and fine-grained stochastic optimization, guided by the principle of greedy optimization. Experimental results demonstrate distinct advantages of the proposed algorithms compared to the baselines. Moreover, transferability verification confirms the applicability of the proposed algorithms across various scenarios and their ability for knowledge transfer.

This paper investigates the problem of decentralized spectrum sharing in self-organizing networks against a dynamic and unknown jamming environment using reinforcement learning. In the network, the anti-jamming spectrum sharing has to not only coordinate spectrum access of users, but also combat the malicious jamming. However, most existing anti-jamming approaches are centralized and require information exchange, which are not suitable for decentralized self-organizing networks in the jamming environment. We formulate the multiuser anti-jamming channel selection problem as a Markov game, and propose a decentralized deep reinforcement learning based collaborative anti-jamming algorithm to achieve the equilibrium solution. It is shown in the simulation part that without information exchange, the approach enables multiple users to independently explore the spectrum environment and obtain effective (close to optimal) collaborative anti-jamming strategies against unknown and dynamic jamming.

Abstract This paper investigates the joint relay and channel selection problem using a deep reinforcement learning (DRL) algorithm for cooperative communications in a dynamic jamming environment. The latest types of jammers include the mobile and smart jammer that contains multiple jamming patterns. This new type of jammer poses serious challenges to reliable communications such as huge environment states, tightly coupled joint action selections and real‐time decision requirements. To cope with these challenges, a DRL‐based relay‐assisted cooperative communication scheme is proposed. In this scheme, the joint selection problem is constructed as a Markov decision process (MDP) and a double deep Q network (DDQN) based anti‐jamming scheme is proposed to address the unknown and dynamic jamming behaviors. Concretely, a joint decision‐making network composed of three sub‐networks is designed and the independent learning method of each sub‐network is proposed. The simulation results show that the user agent is able to anticipate the jammer behaviors and elude the jamming in advance. Furthermore, compared with the sensing‐based algorithm, the Q learning‐based algorithm and the existing DRL‐based anti‐jamming approaches, the proposed algorithm maintains a higher average normalized throughput.

This article investigates the problem of synchronous randomized neighbor discovery with directional antennas. Due to the long tail effect, it will take long time to discover the last few neighbors, which increases overall neighbor discovery time. This effect is due to small proportion of remaining undiscovered neighbors. Moreover, improper choices of reception probabilities make the discovery even worse. In this article, a cognitive framework is proposed to minimize the expectation of neighbor discovery time. We present a scheme in which reception probabilities are dynamically adjusted. We consider an ideal scenario and a practical scenario. In an ideal scenario where perfect information about the number of neighbors is available, reception probabilities are adjusted according to the number of neighbors. A method of dynamic programming is used to recursively calculate the optimal reception probabilities. In an actual scenario where perfect information about number of neighbors is unavailable, a neighbor estimation method based on maximum-likelihood estimation is executed before probability adjustment. Simulation results show that when perfect information about neighbor is available and total transmission probability is within a proper range (between 0.1 and 0.2), the average neighbor discovery time can be significantly reduced (by 38% to 43%, respectively) compared with an existing probability-fixed scheme. With imperfect information, the scheme also works well and realizes appreciable reduction in average neighbor discovery time compared with existing self-adaptive schemes.

This paper investigates the anti-jamming problems in wireless communication networks. In these networks consisted of multiple devices (users), there exist two critical problems. On the one hand, users with various transmission requirements should coordinate their channel selection strategies distributedly to avoid spectrum conflicts and satisfy transmission demands. On the other hand, they also need to fully consider how to eliminate the effects of malicious attacks. To cope with the internal coordination and external confrontation challenges and accommodate the dynamic changing jamming attacks, a context-aware distributed coordinated anti-jamming channel access mechanism is proposed, which means for different cases of jamming attacks, different access strategies are adopted. In detail, to reflect the heterogeneous communication demands of users, the transmission satisfaction function is firstly introduced. Then, the multi-user anti-jamming scenario is modeled as a context-aware multi-pattern dynamic anti-jamming game, which can be decomposed into two sub-games. Here, for the case that the control channel is available, a local altruistic sub-game is introduced. While for the case that the control channel has been jammed, an anti-jamming congestion sub-game is designed. Besides, the existence of Nash Equilibriums is demonstrated. To obtain NEs, a context-aware distributed channel access (CDCA) algorithm is designed. Through game-theoretic analysis and distributed learning, global transmission satisfaction can be improved under the dynamic jamming environment. Furthermore, the fairness of the network can also be guaranteed.

Informal discussions on social platforms (e.g., Stack Overflow) accumulates a large body of programming knowledge in natural language text. Natural language process (NLP) techniques can be exploited to harvest this knowledge base for software engineering tasks. To make an effective use of NLP techniques, consistent vocabulary is essential. Unfortunately, the same concepts are often intentionally or accidentally mentioned in many different morphological forms in informal discussions, such as abbreviations, synonyms and misspellings. Existing techniques to deal with such morphological forms are either designed for general English or predominantly rely on domain-specific lexical rules. A thesaurus of software-specific terms and commonly-used morphological forms is desirable for normalizing software engineering text, but very difficult to build manually. In this work, we propose an automatic approach to build such a thesaurus. Our approach identifies software-specific terms by contrasting software-specific and general corpuses, and infers morphological forms of software-specific terms by combining distributed word semantics, domain-specific lexical rules and transformations, and graph analysis of morphological relations. We evaluate the coverage and accuracy of the resulting thesaurus against community-curated lists of software-specific terms, abbreviations and synonyms. We also manually examine the correctness of the identified abbreviations and synonyms in our thesaurus. We demonstrate the usefulness of our thesaurus in a case study of normalizing questions from Stack Overflow and CodeProject.

In this paper, we propose a novel method for single-channel blind separation of nonoverlapped sources and, to the best of our knowledge, apply it for the first time to automatic segmentation of lung tumors in positron emission tomography (PET) images. Our approach first converts a 3-D PET image into a pseudo-multichannel image. Afterward, regularization free sparseness constrained non-negative matrix factorization is used to separate tumor from other tissues. By using complexity based criterion, we select tumor component as the one with minimal complexity. We have compared the proposed method with threshold based on 40% and 50% maximum standardized uptake value (SUV), graph cuts (GC), random walks (RW), and affinity propagation (AP) algorithms on 18 nonsmall cell lung cancer datasets with respect to ground truth (GT) provided by two radiologists. Dice similarity coefficient averaged with respect to two GTs is: 0.78 ± 0.12 by the proposed algorithm, 0.78 ± 0.1 by GC, 0.77 ± 0.13 by AP, 0.77 ± 0.07 by RW, and 0.75 ± 0.13 by 50% maximum SUV threshold. Since the proposed method achieved performance comparable with interactive methods, considering the unique challenges of lung tumor segmentation from PET images, our findings support possibility of using our fully automated method in routine clinics. The source codes will be available at www.mipav.net/English/research/research.html.

In this paper, a fully automatic method is proposed to segment the kidney into multiple components: renal cortex, renal column, renal medulla and renal pelvis, in clinical 3D CT abdominal images. The proposed fast automatic segmentation method of kidney consists of two main parts: localization of renal cortex and segmentation of kidney components. In the localization of renal cortex phase, a method which fully combines 3D Generalized Hough Transform (GHT) and 3D Active Appearance Models (AAM) is applied to localize the renal cortex. In the segmentation of kidney components phase, a modified Random Forests (RF) method is proposed to segment the kidney into four components based on the result from localization phase. During the implementation, a multithreading technology is applied to speed up the segmentation process. The proposed method was evaluated on a clinical abdomen CT data set, including 37 contrast-enhanced volume data using leave-one-out strategy. The overall true-positive volume fraction and false-positive volume fraction were 93.15%, 0.37% for renal cortex segmentation; 83.09%, 0.97% for renal column segmentation; 81.92%, 0.55% for renal medulla segmentation; and 80.28%, 0.30% for renal pelvis segmentation, respectively. The average computational time of segmenting kidney into four components took 20 seconds.