Goal-Oriented Middleware Filtering at Transport Layer Based on Value of Updates

TL;DR

This work addresses real-time networked control under constrained resources by introducing a goal-oriented transport-layer middleware. The VoU-based TL framework uses a Belief Network, Augmentation, and a Delay Predictor to admit status updates whose expected impact on control performance outweighs transmission costs, adapting to network congestion. Experimental results on Industrial IoT hardware and Internet-scale setups show substantial improvements in LQG cost (up to ~60% over naive GO TL) and successful stabilization of multiple control loops, highlighting the practical viability of GO semantics at the transport layer. The approach demonstrates how cross-layer, theory-grounded admission decisions can significantly reduce network load while preserving or enhancing control performance in heterogeneous networks.

Abstract

This work explores employing the concept of goal-oriented (GO) semantic communication for real-time monitoring and control. Generally, GO communication advocates for the deep integration of application targets into the network design. We consider CPS and IoT applications where sensors generate a tremendous amount of network traffic toward monitors or controllers. Here, the practical introduction of GO communication must address several challenges. These include stringent timing requirements, challenging network setups, and limited computing and communication capabilities of the devices involved. Moreover, real-life CPS deployments often rely on heterogeneous communication standards prompted by specific hardware. To address these issues, we introduce a middleware design of a GO distributed Transport Layer (TL) framework for control applications. It offers end-to-end performance improvements for diverse setups and transmitting hardware. The proposed TL protocol evaluates the Value of sampled state Updates (VoU) for the application goal. It decides whether to admit or discard the corresponding packets, thus offloading the network. VoU captures the contribution of utilizing the updates at the receiver into the application's performance. We introduce a belief network and the augmentation procedure used by the sensor to predict the evolution of the control process, including possible delays and losses of status updates in the network. The prediction is made either using a control model dynamics or a Long-Short Term Memory neural network approach. We test the performance of the proposed TL in the experimental framework using Industrial IoT Zolertia ReMote sensors. We show that while existing approaches fail to deliver sufficient control performance, our VoU-based TL scheme ensures stability and performs $\sim$$60\%$ better than the naive GO TL we proposed in our previous work.

Figures (10)

• Figure 1: The considered scenario with multiple control loops sharing the network. Being a middleware approach, the proposed GO TL resides between the sensor's network and application processes. If the value of the newly generated state update ($VoU_k$) is positive, the update is admitted to underlying networking TL (e.g., UDP). Value is defined as a difference between the relevance $R_k$ w.r.t. application and the transmission cost $C_k$, the details of which are given in Section \ref{['sec:vou_tl']}.
• Figure 2: The construction of BN, with possible combinations of the OPs' network states and their probabilities estimated based on collected ACKs statistics. The example with $2$ OPs is given.
• Figure 3: The schemes representing AUGM procedure. The controller observation history is augmented by combining the statistics of ACKed packets and the states defined in nodes of BN. The estimation is inferred from the observation history by repeating the controller process. Finally, the summation over all the nodes with the weights (probabilities) given by BN is done.
• Figure 4: Structure of the LSTM model. The NN uses historical states and the tags to predict the next control input $\widetilde{\bm{u}}^{n, lstm}_{k+1}$ and the next state. The procedure is repeated $T_{pr}$ times, with predicted states used as input for subsequent inferences.
• Figure 5: The Delay Predictor scheme with example delay prediction for 2-hop network. The predictor fits the curve into $(IST_t, DELAY_t)$ samples. The current update's expected delay is normalized and scaled with $\lambda$ to get the transmission cost $C_k$.