Deep learning models, albeit having remarkable performance in various image and vision tasks, are notoriously parameter-inefficient. The huge number of parameters in these models makes them unsuitable for large-scale deployment on resource-constrained devices due to high memory and computational requirements. It has also been observed in the literature that many parameters in such deep models are redundant. Hence, there is a need to extract parameter-efficient deep models for wide-scale deployment on resource-constrained devices. To address this challenge, we propose a novel parameter reduction framework using Reinforcement Learning (RL) that performs layer-wise structured pruning of deep models. The proposed method leverages the Proximal Policy Optimization (PPO) strategy to assign learnable importance scores to structural units like filters and neurons across all layers of a trained deep model. These scores get sampled from a Bernoulli distribution to determine the pruning decisions. The PPO agents learn using accuracy-aware rewards from a downstream classification task. Additionally, we propose a compression slider mechanism that enables user-defined control over compression rates, ensuring balanced sparsity. Experiments on benchmark models such as VGG16 and AlexNet across two datasets demonstrate that our method achieves better trade-offs between model performance and compression, while offering finer control over pruning granularity. For VGG16 over CIFAR-10, we achieve a total compression of 85.14% with a drop of 2.08% in inference accuracy. Also, for AlexNet, we prune the model by 77.67%, while observing a drop of 2.84% in accuracy. Moreover, our proposed approach enables precise control over the pruning process, allowing for flexible and targeted compression strategies suitable for real-world deployment.

Rapid urbanization across the cities of the world, led by unprecedented population growth and infrastructure development, has led to significant expansion into the peri-urban areas. This expansion is accompanied by dynamic socio-spatial transformations posing pertinent challenges to environmental sustainability. The Greater Melbourne region, Australia's largest city, has witnessed tremendous growth over the last decade, leading to its expansion toward its peripheries. However, there have been limited studies on the nature of this urban expansion. Our study uses a combination of spatial and socio-economic parameters to develop a novel methodology comprising K-means clustering and Artificial Neural Network (ANN) to demarcate urban expansion around metropolitan Melbourne. Our results revealed a significant increase in the urban regions from 6.06% in 2013 to 9.84% in 2020. Whereas the Green wedge has declined rapidly during the same period signifying expansion of urban into the regions hitherto used for non-urban purposes. Our approach integrating spatial and socio-economic indicators provides a comprehensive understanding of the metropolitan expansion, facilitating the delineation of urban boundaries and planning effective urban management strategies.

The peripheries of metropolitans are undergoing rapid socio-spatial changes due to economic transformations, especially in the global south, leading to the emergence of peri-urban areas. However, there are limited studies analyzing peri-urban expansion in the past decade in countries like India. This could be attributed to the absence of socio-economic datasets that are generally derived from census-related exercises. In this context, our study introduces a machine learning-based model to demarcate peri-urban areas around Hyderabad, India. Our model integrates multiple spatial parameters and domain knowledge driven thresholds using a Radial Basis Function Support Vector Machine (RBF-SVM). Notably, the findings reveal a remarkable 107.96% increase in peri-urban regions around Hyderabad, India between 2013 and 2020. Furthermore, the study performs a robustness analysis, assessing the model’s sensitivity to threshold adjustments. It is important to underscore that the proposed model holds relevance for developing countries, where the timely availability and accuracy of socio-economic data pose notable challenges for peri-urban classification.

Providing a high Quality of Experience (QoE) for video streaming in 5G and beyond 5G(B5G) networks is challenging due to the dynamic nature of the underlying network conditions. Several Adaptive Bit Rate (ABR) algorithms have been developed to improve QoE, but most of them are designed based on fixed rules and unsuitable for a wide range of network conditions. Recently, Deep Reinforcement Learning (DRL) based Asynchronous Advantage Actor-Critic (A3C) methods have recently demonstrated promise in their ability to generalise to diverse network conditions, but they still have limitations. One specific issue with A3C methods is the lag between each actor’s behavior policy and central learner’s target policy. Consequently, suboptimal updates emerge when the behavior and target policies become out of synchronization. In this paper, we address the problems faced by vanilla-A3C by integrating the on-policybased multi-agent DRL method into the existing video streaming framework. Specifically, we propose a novel system for ABR generation- Proximal Policy optimization-based DRL for Adaptive Bit Rate streaming (PPO-ABR). Our proposed method improves the overall video QoE by maximizing sample efficiency using a clipped probability ratio between the new and the old policies on multiple epochs of minibatch updates. The experiments on real network traces demonstrate that PPO-ABR outperforms stateof-the-art methods for different QoE variants.

Adaptive bitrate (ABR) algorithms are used to adapt the video bitrate based on the network conditions to improve the overall video quality of experience (QoE). Recently, reinforcement learning (RL) and asynchronous advantage actor-critic (A3C) methods have been used to generate adaptive bit rate algorithms and they have been shown to improve the overall QoE as compared to fixed rule ABR algorithms. However, a common issue in the A3C methods is the lag between behaviour policy and target policy. As a result, the behaviour and the target policies are no longer synchronized which results in suboptimal updates. In this work, we present ALISA: An Actor-Learner Architecture with Importance Sampling for efficient learning in ABR algorithms. ALISA incorporates importance sampling weights to give more weightage to relevant experience to address the lag issues with the existing A3C methods. We present the design and implementation of ALISA, and compare its performance to state-of-the-art video rate adaptation algorithms including vanilla A3C implemented in the Pensieve framework and other fixed-rule schedulers like BB, BOLA, and RB. Our results show that ALISA improves average QoE by up to 25%-48% higher average QoE than Pensieve, and even more when compared to fixed-rule schedulers.

This study uses data from January 2020 to July 2022 to develop an artificial neural network (ANN)-based model to retrieve atmospheric water vapour pressure profiles (Wv) from COSMIC-2 (Constellation Observing System for Meteorology, Ionosphere) observations over the Indian region. Prior to using the COSMIC-2 (C2) refractivity, we evaluated the C2 refractivity’s quality by comparing it with radiosonde (RS) and ERA5 (ECMWF Reanalysis v5) data. The average and standard deviation of the differences between the refractivity of C2 and that of ERA5 and radiosondes are, respectively, within ±0.5% and less than 2% in the middle and upper troposphere, but there are larger (3 to 5%) differences in the lower troposphere. The C2 refractivity, latitude, longitude, height, month, as well as the knowledge of whether a specific pixel is over land or water are all inputs to ANN model, and its output consists Wv. The operationally retrieved 1D-Var (One-dimensional variational) based Wv from UCAR(University Corporation for Atmospheric Research)and NESDIS (National Environmental Satellite, Data, and Information Service), as well as the retrieved Wv using the proposed ANN model, are compared with ERA5 and RS. Overall, the proposed model outperforms NESDIS retrieved Wv, while the Wv of the proposed models have less error than those of UCAR in the lower atmosphere and perform comparable to UCAR in the middle and upper troposphere. Given that the proposed model can retrieve the Wv solely from C2 data and does not require any external meteorological data, the results are quite promising.

Increasing water pollution necessitates frequent monitoring of water bodies, which is quite time-consuming and costly when done manually, especially for remote locations. Multi-rotor drones can be used for collecting water samples and transporting them to the laboratory for testing, since they are convenient for carrying small payloads and offer increased accessibility compared to aquatic vehicles. Most of the solutions proposed so far have focused on designing a specialized drone for this purpose, which makes the solution extremely difficult to replicate and use. In this paper, we abstract the sampling and water quality testing functionalities to a specialized module which can be attached to any drone. The attachment has its own power supply and does not consume power from the carrier drone. The practicality of the proposed approach is validated by performing a flight mission, including sample collection and water quality analysis using on board sensors. Temperature, pH and electrical conductivity values sensed from the water samples collected using DAWSSM and by manual collection are compared and found to have no difference. This demonstrates the efficacy of DAWSSM as a substitute for the conventional process.

This paper presents NANCY, a system that generates adaptive bit rates (ABR) for video and adaptive network coding rates (ANCR) using reinforcement learning (RL) for video distribution over wireless networks. NANCY trains a neural network model with rewards formulated as quality of experience (QoE) metrics. It performs joint optimization in order to select: (i) adaptive bit rates for future video chunks to counter variations in available bandwidth and (ii) adaptive network coding rates to encode the video chunk slices to counter packet losses in wireless networks. We present the design and implementation of NANCY, and evaluate its performance compared to state-of-the-art video rate adaptation algorithms including Pensieve and robustMPC. Our results show that NANCY provides 29.91% and 60.34% higher average QoE than Pensieve and robustMPC, respectively.

The extreme growth and adoption of Social Media, in combination with their poor governance and the lack of quality control over the digital content being published and shared, has led information veracity to a continuous deterioration. Current approaches entrust content verification to a single centralised authority, lack resilience towards attempts to successfully "game" verification checks, and make content verification difficult to access and use. In response, our ambition is to create an open, democratic, pluralistic and distributed ecosystem that allows easy access to various verification services (both internal and third-party), ensuring scalability and establishing trust in a completely decentralized environment. In fact, this is the ambition of the EU H2020 SocialTruth project. In this paper, we present the innovative project approach and the vision of effective online disinformation detection for various practical use-cases.

Carbon Monoxide (CO) induced by traffic pollution is highly dynamic and non-linear. In a pilot research, we collected some fine-grained 1Hz CO pollution data from a residential road and a busy motorway in Hyderabad, India, in preparation of the deployment of a larger scale, longer term wireless sensor monitoring system. Power conservation is an important issue as the sensor nodes are battery operated. We studied the characteristics of the collected data and designed an adaptive sampling algorithm, Bi-Scale temporal sampler, which adapts the sampling frequency to the statistics collected in real time. This design has incorporated practical engineering considerations including minimising electronic noise, sensor warm-up time and data characteristics. Results show that Bi-Scale sampler achieves better energy saving and statistical deviation ratio for our requirements than burst sampling and eSENSE sampling strategies, which are techniques popularly used in environmental monitoring applications.

We propose a real-time adaptive- spatial sampling technique for the efficient collection of fine grained data in wireless sensor networks. The collection of fine grained data can incur high energy costs. This energy costs can be reduced by exploiting the spatial correlations of adjacent nodes, where only the most dominant nodes collect the data. We show that, using concepts developed in Random Matrix Theory, it is possible to determine the dominant nodes which enable to process noisy data in a time efficient, scalable, decentralized manner. The proposed technique has been validated using spatially interpolated pollution datasets giving good results in terms of data reduction and accuracy.

Air pollution data exhibit characteristics like long range correlations and multi-fractal scaling that can be exploited to implement an energy efficient, adaptive spatial sampling technique for pollution sensor nodes. In this work, we present a) results from de-trended fluctuation analysis to prove the presence of non-linear dynamics in real pollution datasets gathered from trials carried out in Cyprus, b) a novel Multi-scale Nearest Neighbors based Adaptive Spatial Sampling (MNNASS) technique that determines the predictability and in turn the directional influences between data from different sensor nodes, and c) performance analysis of the algorithm in terms of energy savings and measurement accuracy.

A pollution data collection exercise was conducted in Hyderabad, India in February, 2012. Fifteen bespoke Carbon Monoxide (CO) monitors were deployed across a small section of a busy highway to collect data for an urban environmental engineering study targeting traffic-generated pollution. The monitors were used to record CO concentrations and temperature; however, in spite of the fact that the monitors were calibrated in advance of deployment, interpretation of the data collected has proved to be challenging. This paper reports the findings of the experiment and proposes a bias-adjustment separation technique that provides a consensual baseline for all the monitors. The result is that spatial variation in the distribution of CO can be studied at snapshots of time. Moreover, the cross-correlations between sensors can be reliability extracted after the bias adjustment.

We present the design of a novel adaptive sampling technique called Exponential Double Smoothing-based Adaptive Sampling (EDSAS), in which the temporal data correlations provide an indication of the prevailing environmental conditions and are used to adapt the sensing rate of a sensor node. EDSAS uses irregular data series prediction to reduce sampling rate in combination with change detection to maintain data fidelity. The prediction method employs Wright's extension to Holt's method of Exponential Double Sampling (EDS) coupled with a change detection mechanism based on exponentially weighted moving averages (EWMA). The main advantages of EDSAS are that it does not require heavy computation, incurs low memory and communication overhead and the prediction model can be implemented with ease on resource constrained sensor nodes. EDSAS has been evaluated by using real urban road traffic Carbon Monoxide (CO) pollution datasets and has been compared and shown to give better results for performance metrics like sampling fraction and miss ratio. We have also undertaken analysis of the pollution data based on the information received and shown that EDSAS scores over other published technique called e-Sense in capturing the underlying characteristics of the real data.