To Compute or not to Compute? Adaptive Smart Sensing in Resource-Constrained Edge Computing

TL;DR

This work tackles adaptive sensing in resource-constrained edge networks where each sensor can transmit raw data or perform local processing, creating a latency-accuracy trade-off for global state estimation. It introduces an estimation-theoretic model that incorporates both computation and communication delays and proposes a reinforcement-learning-based method to allocate sensing resources online, including a sleep mode to reduce processing load. The approach is validated in two scenarios—drone-based target tracking and autonomous driving—showing that learned online sensor selection can reduce estimation error and, in some cases, significantly reduce energy consumption compared to static designs. By marrying model-based Kalman estimation with data-driven policy optimization, the paper demonstrates a practical, scalable path to adaptive, latency-aware sensing in edge-enabled cyber-physical systems.

Abstract

We consider a network of smart sensors for an edge computing application that sample a time-varying signal and send updates to a base station for remote global monitoring. Sensors are equipped with sensing and compute, and can either send raw data or process them on-board before transmission. Limited hardware resources at the edge generate a fundamental latency-accuracy trade-off: raw measurements are inaccurate but timely, whereas accurate processed updates are available after processing delay. Hence, one needs to decide when sensors should transmit raw measurements or rely on local processing to maximize network monitoring performance. To tackle this sensing design problem, we model an estimation-theoretic optimization framework that embeds both computation and communication latency, and propose a Reinforcement Learning-based approach that dynamically allocates computational resources at each sensor. Effectiveness of our proposed approach is validated through numerical experiments motivated by smart sensing for the Internet of Drones and self-driving vehicles. In particular, we show that, under constrained computation at the base station, monitoring performance can be further improved by an online sensor selection.

Figures (16)

• Figure 1: Scheme of the proposed methodological framework: the RL algorithm learns a sensing design to maximize performance of the estimation algorithm.
• Figure 2: Data collection and transmission. Computation at the $i$th sensor is ruled by sensing policy $\pi_{i}$. Here, sensing decisions $\{\gamma_{k_j}^{i}\}_{j=0}^3 = \{\mathrm{p},\mathrm{p},\mathrm{s},\mathrm{r}\}$ are shown and \ref{['eq:sampling-sequence-subequations']} reads $\mathcal{K}_{i}[0] = s_i^0(k_0) = k_0$, $\mathcal{K}_{i}[1] = s_i(k_0) = k_1$, and $\mathcal{K}_{i}[2] = s_i(k_1) = k_3$. Measurements are received after delays induced by local computation (rectangular blocks) and communication (dashed arrows). For example, under $\gamma_{k_0}^{i} = \mathrm{p}$, the sample acquired at time $k_0$ is first processed (with processing delay $\tau_ {{i,\text{proc}}}$), then transmitted at time $k_1 = k_0 + \tau_ {{i,\text{proc}}}$ (with communication delay $\delta_ {{i,\text{proc}}}$), and finally received at the base station at time $k_1 + \delta_ {{i,\text{proc}}} = k_0 + \Delta_ {{i,\text{proc}}}$ (with delay at reception $\Delta_ {{i,\text{proc}}}$).
• Figure 3: Data processing at the base station. Resource-constrained centralized processing introduces fusion delay$\phi_{k_5}$ to estimate $x_{k_5}{}$. Measurements $y_{k_1}^{(i)}$ and $y_{k_3}^{(j)}$ are received before computation starts at time $k_4 = k_5 - \phi_{k_5}$ and are used to compute $\hat{x}_{k_5}^{}$, i.e.,$y_{k_1}^{(i)}, y_{k_3}^{(j)} \in\mathcal{Y}_{k_5}^{}$, while $y_{k_2}^{(l)}$ is received after time $k_4$ and cannot be used in estimation of $x_{k_5}$, i.e.,$y_{k_2}^{(l)} \notin\mathcal{Y}_{k_5}^{}$.
• Figure 4: Real-time estimation at the base station. The state estimate is updated at each point in time (top). Because of limited resources at the base station, open-loop updates are performed whenever fresh sensory data are being processed (bottom), causing estimation to degrade overtime through additive noise $w_{k}$ in nominal dynamics \ref{['eq:stateEquation']}. As soon as the data processing subroutine produces an updated estimate with new measurements, e.g.,$\hat{x}_{k_1}^{}$ at time $k_1$, the estimation inaccuracy is reduced. Note that the top plot is qualitative: the estimate quality does not degrade linearly, in general.
• Figure 5: Homogeneous sensing policy. Sampling and data processing at identical sensors are ruled by policy $\pi_ {{\text{hom},empty}}$. Decision $\gamma_{\ell}^ {\text{hom},empty}$ is communicated at time $k^{(\ell)}_{}$ and realized at individual sensors as $\gamma_{\ell}^{i} = \mathrm{r}$ and $\gamma_{\ell}^{j} = \mathrm{s}$. Concurrently, the $i$th sensor disregards its current processed measurement (red cross) and switches to raw mode, acquiring a new sample at time $k^{(\ell)}_{}$.

Theorems & Definitions (10)

• Remark 1: Sensor processing
• Definition 1: Sensing policy
• Remark 2: Real-time estimation
• Remark 3: Impact of processing on estimation
• Remark 4: Novelty of sensor selection
• Definition 2: Homogeneous sensing policy
• Definition 3: Network sensing policy
• Remark 5: System dynamics and computational complexity
• Remark 6: Energy saving
• Remark 7: Computational scalability