HARL enables agents to execute cooperative tasks by adopting agent-specific policies. Most of existing HARL methods use individual policy neural networks to ensure monotonic improvement, which leads to substantial computational overhead. The proposed SDE-HARL overcomes this limitation by decomposing each agent's policy neural network into a lightweight local neural network and a global neural network executed at an edge server. Each local neural network generates and sends a compressed latent representation to the edge server, which aggregates the representations and produces agent-specific inferences. As such, SDE-HARL allows to significantly save computing and networking resources while preserving agent-specific behavior. A key feature of SDE-HARL is grouping agents with similar roles via a role-aware mechanism and share partial parameters in their global networks, while an identity-aware mechanism is introduced to promote behavioral diversity among agents within the same group. We prototyped SDE-HARL on an experimental testbed composed of a Jetson Nano and Raspberry PI to measure latency and network resource consumption. We evaluated SDE-HARL's performance on several benchmark datasets, including Google Research Football and StarCraft II. Experimental results show that SDE-HARL reaches up to 90% win rate while reducing latency, energy consumption, and networking overhead respectively by 2x, 2.5x, and 5x compared to existing work.

Internet of Things (IoT) devices need to continuously adapt their wireless receiver based on the varying communication conditions to achieve the right trade-off between system-level performance and energy efficiency. Even though the environmentadaptability functions are developed in the back end of receiver chains, existing front-ends are designed for a fixed communication standard. The only adaptive wireless systems equipped with tunable front-ends use traditional optimization methods to reconfigure their circuit. The first issue is that these methods require a significant time to converge to the optimal configuration. The second limitation is that they cannot generalize to unseen propagation environments and operational conditions. In this paper, we propose Front-End Adaptation with Reinforcement Learning (FEARL) to dynamically optimize front-end circuits considering the ongoing distortion and interference levels with the aim to achieve optimal wireless system-level performance. FEARL characterizes the wireless link quality using the received baseband samples and reconfigures the front-end in real-time to realize an end-to-end optimized energy-efficient receiver. We developed a framework with the circuit simulator-in-the-loop, which is further utilized to train and evaluate the FEARL policy. The results show that FEARL is$7.8 x$times faster in responding to variations in the wireless environment, and is able to find a satisfactory circuit configuration even in unseen conditions.

Reconfigurability and self-optimization have become essential during radio frequency integrated circuit (RFIC) design to support the growing number of devices and fast changes in the surrounding wireless spectrum. This has created the need to develop new design approaches for RFICs based on end-to-end wireless system-level performance metrics during operation in dynamically changing communication environments. This paper introduces a CMOS mixer with a wide range of digitally tunable bias current for machine learning (ML) based adaptation. The mixer is designed to become part of a self-adaptive receiver (RX) architecture that is capable of optimizing its performance in accordance to wireless channel conditions by evaluating systemlevel parameters. The proposed mixer topology has digitally tunable bias current and a programmable helper current (PHC) circuit to maintain the voltage headroom during operation with a wide tuning range. It is capable of dynamically minimizing power consumption based on performance and wireless network requirements. Post-layout simulation results show that the power consumption can be reduced up to 8x depending on momentary performance needs, while maintaining an IIP3 greater than -2.4 dBm for the entire range of operation.