Multi-modal reinforcement learning (RL) has been brought into focus due to its ability to provide complementary information from different sensors, enriching observations of agents. However, the introduction of multi-modal highdimensional observations brings challenges to sample efficiency. There is a lack of research on how to efficiently obtain multi-modal latent states while encouraging them to generate complementary information. To address this, we propose a representation learning method, Multi-modal Joint Predictive Representation (MJPR), which utilizes multi-modal interactive information to predict future latent states. The joint prediction method achieves the representation training for modalities and promotes each modality to generate complementary information related to predictions of each other. In addition, we introduce multi-modal loss balancing to prompt training equilibrium and cross-modal contrastive learning (CMCL) to align the modalities for effective modal interaction. We establish the multi-modal environments in the Deepmind Control suite (DMC) and Webots and compare our method with current RL representation methods. Experimental results show that MJPR outperforms state-of-the-art methods by an average of 12.0% on six subtasks in DMC environments. It outperforms advanced methods by 16.7% and 55.4% in simple tasks and complex tasks of Webots environment, respectively. Moreover, ablation experiments are established in the DMC environment to verify the importance of each module to MJPR.

The Society of Automotive Engineers (SAE) J2954 Differential Inductive Positioning System (DIPS) is a ground breaking technology introduced to enable positioning technique for electric vehicle (EV) wireless power transfer (WPT). This technology and its related standard have great potential to bring EV wireless charging to mass production and opens the doors for commercializing autonomous vehicles. Although the DIPS standard has defined the hardware requirements [1], the positioning algorithm design has not been addressed in the literature. In this paper, we introduce the industry's first algorithm for DIPS. We mathematically derive the signal model and parameters estimation algorithm, then evaluate the estimation accuracy of the proposed algorithm using Monte Carlo simulations. The evaluation results have shown the algorithm can achieve centimeter-level accuracy and approach the Cramér-Rao bound.

Effective skill learning in an unsupervised manner is one of the capabilities an intelligent agent or robot should have. The discovered task-agnostic skills can be fine-tuned to downstream long-horizon tasks to improve execution efficiency. Unfortunately, the self-learning of locomotion skills, which occurs naturally in infancy, has been slow to develop in robotics. The instability exhibited by existing skill-learning methods makes it difficult to directly apply to complex control tasks, such as humanoid robots. To acquire reliable robotic locomotion skills, this article proposes a controllable latent space partition framework to assist reinforcement learning in accomplishing practicability-oriented unsupervised skill discovery (PoSD). Specifically, we use the distance similarity measure of the trajectory feature space to introduce the indicative information of the expert demonstrations into the partitioning and mapping process of the latent space. In addition, the intrinsic subrewards based on contrastive learning and particle entropy are designed to promote skill diversity and encourage exploration. Finally, reinforcement learning completes the generation of skill-conditioned policy driven by composite intrinsic rewards. The performance investigation of our method is conducted on five robots with more than 15 skills. The results indicate that PoSD achieves noticeable improvements in adaptation efficiency and practicability compared with other SOTA unsupervised skill discovery methods.

This innovative practice full paper proposes Creator-Thon, a novel teaching project designed to introduce Generative AI into traditional interior design workflows. The project adopts an interdisciplinary approach including application scenario investigation, workflow construction, collaborative innovation, and iterative optimization. The underlying research is grounded in a six-level goal hierarchy, comprising sub-goals of integrating AI capabilities and design knowledge, embedding AI skills into workflows, enhancing the learning engagement, empowering personal career development, fostering (human-AI and human-human) collaboration, and transforming the design paradigms. The teaching framework is progressively refined across multiple iterations, enhancing learning engagement by simplifying tool operations, refining design language, and re-engineering the design process. The findings underscore that integration of authentic application scenarios markedly boosts engagement and motivation among learners, modular instruction also enhances teaching quality and practical skills, and the adaptive optimization of the teaching framework informed by participants' feedback, is key to a successful Generative AI involved interior design pedagogy. This work provides insights that contribute to the application of Generative AI technology in interior design education and practice, facilitating transformations and progress in AI literacy in the field.

A scarcity of frequencies arises from the increased demand for the Industrial Internet of Things (IIoT) and networked systems in warehouse operations. This puts an additional burden on available bandwidth in cellular networks. This problem can be tackled by cognitive radio networks (CRNs), which use spectrum sensing to track and access unused frequencies, increasing spectrum efficiency. However CRNs have been extensively studied in several IoT projects, this study is the initial to examine the utilisation of CRN to intelligently manage the use of available radio frequencies. Two macro base stations are the central hubs of a network, communicating with IIoT devices in warehouse settings. This paper investigates the utilisation of CRN with machine-learning algorithms to intelligently manage the spectrum for connected IoT devices for intelligent Warehouse settings. A range of machine learning methods, including Support Vector Machine (SVM), k-nearest Neighbors (KNN), Decision Tree, Random Forest, and Naive Bayes, are provided to identify accessible bands for the best possible spectrum allocation and to recognize key users. Based on criteria like accuracy, precision, recall, and score for all ML techniques, the system’s performance is assessed. The numerical results demonstrate a noteworthy 20% reduction in false positives and a substantial improvement in cooperative spectrum sensing accuracy, which in turn improves the effectiveness of IIoT operations in warehouse environments.

With the rise of requirements for high-speed and low-latency connectivity, innovative approaches such as network slicing, Quality of Service (QoS) Provisioning, and reinforcement learning-based resource allocation, including the use of resource blocks (RBs) including radio resources, have to keep pace with these evolving requirements. By utilising network intelligence through machine learning and deep reinforcement learning, there is a potential to enhance QoS provisioning, expand the current network capacity, reduce congestion and latency, improve energy efficiency, and thus support new business models and revenue sources. The progress so far suggests that while considering RBs, network-slicing datasets, intertwined with QoS provisioning, have not been explored comprehensively, and packet drop probability or rate has not been taken into account in resource allocation. This paper proposes a network slicing method that uses seven machine learning algorithms and demonstrates its efficiency and accuracy compared to benchmarks in the literature with respect to QoS. Moreover, a priority algorithm is developed to ensure that packets with a high chance of being dropped (affecting QoS) are queued first. A resource allocation algorithm considering QoS provisioning and RBs is utilised to improve network performance based on a mathematical derivation of packet drop rate. Furthermore, a virtualisation of the processing between Cloud and Edge depends on the network slice. By intelligently distributing tasks between Cloud and Edge resources using deep reinforcement learning and genetic algorithms, an offloading script ensures uninterrupted service availability even when all network resources (i.e., RBs) are in use, thus maintaining the desired QoS.

Federated learning enables a large number of clients (such as edge computing devices) to learn a model jointly without data sharing. However, the high amount of communication of the federated learning aggregation algorithm hinders the realisation of artificial intelligence in the last mile. Although FederatedAveraging (FedAvg) is the leading algorithm, its communication cost is still high. The method of communication delay and gradient sparse can reduce the communication cost, but there is no previous work to analyse the relationship and common effects of these two dimensions. Aiming at the problems that federated learning communication is expensive and it has become a training bottleneck, we improve the FedAvg algorithm and propose an adaptive communication frequency FederatedAveraging algorithm (AFedAvg). The gradient sparse operation in the algorithm reduces the quantity of parameters for a single communication, while the communication delay operation allows training to converge faster and obtain smaller losses. The number of sparse parameters is used to select the communication frequency of next round dynamically. Experimental results prove that, the AFedAvg algorithm is superior to the FedAvg and its variants in terms of communication cost. It achieves 2.4X–23.1X communication compression in different data distributions with minimal communication rounds required by the algorithm to converge.

This paper studies the affine frequency division multiplexing (AFDM)-empowered sparse code multiple access (SCMA) system, referred to as AFDM-SCMA, for supporting massive connectivity in high-mobility environments. First, by placing the sparse codewords on the AFDM chirp subcarriers, the input-output (I/O) relation of AFDM-SCMA systems is presented. Next, we delve into the generalized receiver design, chirp rate selection, and error rate performance of the proposed AFDM-SCMA. The proposed AFDM-SCMA is shown to provide a general framework and subsume the existing OFDM-SCMA as a special case. Third, for efficient transceiver design, we further propose a class of sparse codebooks for simplifying the I/O relation, referred to as I/O relation-inspired codebook design in this paper. Building upon these codebooks, we propose a novel iterative detection and decoding scheme with linear minimum mean square error (LMMSE) estimator for both downlink and uplink channels based on orthogonal approximate message passing principles. Our numerical results demonstrate the superiority of the proposed AFDM-SCMA systems over OFDM-SCMA systems in terms of the error rate performance. We show that the proposed receiver can significantly enhance the error rate performance while reducing the detection complexity.

The objective of this work is to address one of the recent ITU AI/machine learning challenges, specifically on the RF power amplifier (PA) behavioral modelling. This paper presents a novel model namely real-valued time-delay temporal convolutional network (RVTDTCN) to learn the long-term memory effect from PA in 5G base-station transmitters. The proposed model employs dilated convolutions to incorporate information from various time stamps and capture intricate temporal dependencies. This is achieved without the requirement of additional coefficients. In our experiment, the ITU ML5G-PS-007 datasets collected from commercial 5G base-stations are employed for both training and testing purposes. The performance of the learned PA behavior is evaluated using both normalized mean square error (NMSE) and adjacent channel error power ratio (ACEPR). The results show that the RVTDTCN model achieves competitive performance with the state-of-the-art feedforward neural network (FNN) models but reducing the computational complexity by 60%. Moreover, the proposed model outperforms the state-of-the-art convolutional neural network (CNN) model by 2-3dB in NMSE and ACEPR, with a similar complexity.

The paper presents four deep feedforward neural network (FNN) models for digital predistortion of radiofrequency power amplifiers (PAs), namely real-valued time-delay/advance NN (RVTDANN), augmented real-valued time-delay/advance NN (ARVTDANN), residual real-valued time-delay/advance NN (R2TDANN) and augmented residual real-valued time-delay/advance NN (AR2TDANN). An experimental platform is implemented to evaluate their performance in terms of adjacent channel leakage ratio, normalized mean square error and error vector magnitude (EVM). The platform uses an envelope tracking PA excited by 160 MHz IEEE 802.11 WiFi signals. It is observed that the proposed models outperform the state-of-the-art FNN models by up to 3.7 dB in EVM, using similar number of model coefficients. Moreover, the proposed models outperform the state-of-the-art recurrent neural network model by up to 0.7 dB in EVM. When the PA operates close to its saturation point with a strong non-linear effect, the proposed RVTDANN and R2TDANN can almost achieve a transmit EVM of −42 dB. The observation indicates that the proposed models are key enablers for wireless transmitters enjoying high power efficiency and high throughput simultaneously.

This paper presents a novel TOA estimation approach by utilizing multipath parameter estimation algorithm to reconstruct channel frequency response on null subcarriers in OFDM systems. The proposed approach is evaluated using legacy long training field defined in IEEE 802.11ac standard, and is shown by simulations to outperform the state-of-the-arts when signal bandwidth is 20-40 MHz in both 802.11 channel model B and WINNER A1 LOS channel.

This paper presents three novel spoofing attack detection approaches based on indoor wireless channel continuity, including a cross-correlation approach, a machine learning approach based on nearest neighbor and a machine learning approach based on neural network. The proposed approaches are evaluated by lab experiments in a wireless chamber. The experimental data used is the IEEE 802.11ac WiFi channel frequency response estimates from continually captured WiFi packets transmitted by a slowly moving mobile phone embedded with a Samsung Exynos WiFi MODEM. The experimental results show that the proposed neural network approach outperforms the other two approaches in terms of spoofing attack detection.

The paper proposes a deep neural network (DNN) based receiver to outperform the state-of-the-art w/o timing synchronization error in orthogonal frequency-division multiplexing (OFDM) systems. Moreover, the closed-form of a traditional minimum mean square error (MMSE) receiver is derived in the presence of inter-symbol-interference. The derived receiver is used to benchmark the performance of the proposed DNN.

Retail mobile location analytics (MLA) is an emerging location based service (LBS) for business intelligence. State-of-the-art cellular positioning techniques are not suitable for such LBS category since they affect the network design and operation. This paper presents Path Intelligence's Global System for Mobile Communications (GSM) sensor technique which enables retail MLA without affecting the design and operation of GSM networks and mobile stations (MSs). Without prior knowledge of uplink resource allocation, the sensor passively detects a MS and extracts its corresponding signals for positioning purposes.

One of the major challenges of Cellular network based localization techniques is lack of hearability between mobile terminals (MTs) and base stations (BSs), thus the number of available anchors is limited. In order to solve the hearability problem, previous work assume some of the MTs have their location information via Global Positioning System (GPS). These located MT can be utilized to find the location of an un-located MT without GPS receiver. However, its performance is still limited by the number of available located MTs for cooperation. This paper consider a practical scenario that hearability is only possible between a MT and its home BS. Only one located MT together with the home BS are utilized to find the location of the un-located MT. A hybrid cooperative localization approach is proposed to combine time-of-arrival and received signal strength based fingerprinting techniques. It is shown in simulations that the proposed hybrid approach outperform the stand-alone time-of-arrival techniques or received signal strength based fingerprinting techniques in the considered scenario. It is also found that the proposed approach offer better accuracy with larger distance between the located MT and the home BS.

Previous work about cooperative localization in cellular networks usually consider a centralized processor (CP) is available for location estimation. This paper consider cooperative localization in a distributed base station (BS) scenario, where there is no CP, and the distributed BSs are responsible for location estimation. In this scenario, Global Positioning System (GPS) enable mobile terminals (MTs), i.e., located MTs, are employed as reference nodes.Then, several located MTs can help to find the locations of an un-located MT, by estimating the distance between the un-located MT using received signal strength techniques. Two localization approaches are proposed, the first approach requires only one BS to collect all the assistance information for localization and estimate the location. The second approach distribute the location estimation task to several BSs. The communication overhead between distributed BSs are investigated for these two approaches. Moreover, by taking into account the effect of imperfect location knowledge of the located MTs, the accuracy limits of both approaches are derived. The simulation results shows that compared with the first approach, the second approach can reduce the communication overhead between distributed BSs with the paid of accuracy.

This paper consider cooperative localization in cellular networks. In this scenario, several located mobile terminals (MTs) are employed as reference nodes to find the location of an un-located MT. The located MTs sent training sequences in the uplink, then the un-located MT perform distance estimation using received signal strength techniques. The localization accuracy of the un-located MT is characterized in terms of squared position error bound (SPEB). By taking into account the imperfect a priori location knowledge of the located MTs, the SPEB is derived in a closed-form. The closed-form indicate that the effect of the imperfect location knowledge on SPEB is equivalent to the increase of the variance of distance estimation. Moreover, based on the obtained closed-form, a MT selection scheme is proposed to decrease the number of located MTs sending training sequences, thus reduce the training overhead for localization. The simulation results show that the proposed scheme can reduce the training overhead with the paid of accuracy. And with the same training overhead, the accuracy of the proposed scheme is better than that of random selection.

Line-of-sight (LoS) time-of-arrival (ToA) is one of key parameters for network-based localization techniques. Its estimation accuracy depends on the bandwidth of transmitted signals. This paper aims to investigate the LoS-ToA estimation for an interleaved OFDMA based network, where the signal bandwidth is not sufficiently wide to offer the acceptable estimation accuracy. In this situation, an over-sampling approach is proposed to improve the resolution. Moreover, the proposed scheme does not require modification of mobile communications systems. It is found that the proposed approach can significantly improve the ToA estimation performance, and offer adequate accuracy even for the low signal-to-noise ratio (SNR) range.