Adaptive Cooperative Distributed Compressed Sensing for Edge Devices: A Multiagent Deep Reinforcement Learning Approach

摘要

We propose a lightweight and adaptive distributed compressed sensing (DCS) with multi-sensor collaboration based on multiagent deep reinforcement learning (LADICS-MARL). To efficiently acquire data generated by sensor nodes deployed over a wide area and for long periods, we previously proposed a lightweight and adaptive compressed sensing method based on deep learning for edge devices, called LACSLE that changes the compression ratio in real-time according to the data between one sender and one receiver using pre-trained deep learning model. LADICS-MARL is an extension for multiple senders and one receiver and supports DCS through which multiple compressed data are simultaneously reconstructed. Multiple sensor nodes cooperate based on multiagent reinforcement learning to estimate the optimal compression ratio for all senders according to each corresponding data, as well as the transmitted compressed data from other sensor nodes. In addition, a gateway optimizes the combination of groups where some compressed data are reconstructed simultaneously. A performance evaluation using acceleration data from multiple sensor terminals acquired on a bridge suggests that the multiagent-based LADICS-MARL can reconstruct original data from less compressed data compared to the single-agent-based LACSLE under the threshold of reconstruct error.