Efficient slot allocation and transmit-sleep scheduling is an effective access control mechanism for improving communication performance and network lifetime in resourceconstrained wireless networks. In this paper, a decentralized and multi-tier framework is presented for joint slot allocation and transmit-sleep scheduling in wireless network nodes with thin energy budget. The key learning objectives of this architecture are: collision-free transmission scheduling, reducing energy consumption, and improving network performance. This is achieved using a cooperative and decentralized learning behavior of multiple Reinforcement Learning (RL) agents. The resulting architecture provides throughput-sustainable support for data flows while minimizing energy expenditure and sleep-induced packet losses. To achieve this, a concept of Context is introduced to the RL framework in order to capture network traffic dynamics. The resulting Contextual Deep Q-Learning (CDQL) model makes the system adaptive to dynamic and heterogeneous network load. It also improves energy efficiency when compared with the traditional tabular Q-learning-based approaches. The results demonstrate how this framework can be used for prioritizing application-specific requirements, namely, energy saving and communication reliability. The trade-offs among packet drop, energy expenditure, and learning convergence are studied, and an application-specific solution is proposed for managing them. The performance is compared against an existing state-of-the-art scheduling approach. Moreover, an analytical model of the system dynamics is developed and validated using simulation for arbitrary mesh topologies and traffic patterns.

A lightweight on-device liquid consumption estimation system involving an energy-aware machine learning algorithm is developed in this work. This system consists of two separate on-device neural network models that carry out liquid consumption estimation with the result of two tasks: the detection of sip from gestures with which the bottle is handled by its user and the detection of first sips after a bottle refill. This predictive volume estimation framework incorporates a self-correction mechanism that can minimize the error after each bottle fill-up cycle, which makes the system robust to errors from the sip classification module. In this paper, a detailed characterization of sip detection is performed to understand the accuracy-complexity tradeoffs by developing and implementing a variety of different ML models with varying complexities. The maximum energy consumed by the entire framework is around 119 $\mathrm{mJ}$ during a maximum computation time of 300 $\mathsf{\mu }\mathrm{s}$. The energy consumption and computation times of the proposed framework is suitable for implementation in low-power embedded hardware that can be incorporated in consumer grade water bottles.

This paper proposes a Reinforcement Learning (RL) framework for joint transmit-sleep scheduling for multi-hops wireless sensor and IoT networks with energy harvesting. The goal of the framework is to learn situation-specific scheduling policies for reducing energy expenditure, while maintaining network performance in terms of packet delivery ratio and end-to-end delay. The proposed system uses a cooperative RL approach where two learning agents, deployed per node, jointly learn transmit and sleep scheduling strategies to manage network energy budgets. The RL framework also uses a localized learning confidence parameter sharing strategy that allows the nodes to ignore unreliable RL observations. This makes the system scalable for network topologies where the source and destination nodes are separated by many numbers of hops. The learning module is decentralized in that learning is independently carried out at the sensor/IoT nodes without relying on a central learning coordinator. Decentralized learning makes the system computationally efficient, and also avoids energy and communication bandwidth overhead usually imposed by a central coordinator. With simulation studies, the proposed learning-driven protocol is tested and compared against existing known MAC sleep schedulers. The proposed mechanism is validated for different scenarios with heterogeneous topologies, traffic, and energy harvesting conditions.

This paper presents a computationally efficient and distributed speaker diarization framework for networked IoT-style audio devices. The work proposes a Federated Learning model which can identify the participants in a conversation without the requirement of a large audio database for training. An unsupervised online update mechanism is proposed for the Federated Learning model which depends on cosine similarity of speaker embeddings. Moreover, the proposed diarization system solves the problem of speaker change detection via. unsupervised segmentation techniques using Hotelling's t-squared Statistic and Bayesian Information Criterion. In this new approach, speaker change detection is biased around detected quasi-silences, which reduces the severity of the trade-off between the missed detection and false detection rates. Additionally, the computational overhead due to frame-by-frame identification of speakers is reduced via. unsupervised clustering of speech segments. The results demonstrate the effectiveness of the proposed training method in the presence of non-IID speech data. It also shows a considerable improvement in the reduction of false and missed detection at the segmentation stage, while reducing the computational overhead. Improved accuracy and reduced computational cost makes the mechanism suitable for real-time speaker diarization across a distributed IoT audio network.

This paper presents an on-device semi-supervised human activity detection system that can learn and predict human activity patterns in real time. The clinical objective is to monitor and detect the unhealthy sedentary lifestyle of a user. The proposed semi-supervised learning (SSL) framework uses sparsely labelled user activity events acquired from Inertial Measurement Unit sensors installed as wearable devices. The proposed cluster-based learning model in this approach is trained with data from the same target user, thus preserving data privacy while providing personalized activity detection services. Two different cluster labelling strategies, namely, population-based and distance-based strategies, are employed to achieve the desired classification performance. The proposed system is shown to be highly accurate and computationally efficient for different algorithmic parameters, which is relevant in the context of limited computing resources on typical wearable devices. Extensive experimentation and simulation study have been conducted on multi-user human activity data from the public domain in order to analyze the trade-off between classification accuracy and computation complexity of the proposed learning paradigm with different algorithmic hyper-parameters. With 4.17 h of training time for 8000 activity episodes, the proposed SSL approach consumes at most 20 KB of CPU memory space, while providing a maximum accuracy of 90% and 100% classification rates.