On the trade-off between event-based and periodic state estimation under bandwidth constraints

Dominik Baumann, Thomas B. Schön

In the world of networked control systems—think swarms of drones or automated factories—bandwidth is the ultimate currency. To keep these systems stable, sensors must send state updates to a remote estimator. The traditional way to do this is "periodic": every agent takes a turn in a fixed order. However, "event-based" methods have gained popularity by promising to transmit data only when "necessary" (e.g., when an error exceeds a threshold).

But there is a catch. Event-based methods aren't free. They require a scheduler to decide who gets to talk, and that scheduling process itself consumes bandwidth. In a new paper, Dominik Baumann and Thomas B. Schön investigate a fundamental trade-off: Is the "intelligence" of event-based scheduling worth the bandwidth it steals from actual data transmission?

TL;DR

The researchers derive a mathematical criterion to determine when "predictive triggering" (a smart scheduling method) outperforms simple periodic communication. They demonstrate that while event-based methods are superior for heterogeneous systems or those facing localized disturbances, simple periodic schedules are often better for uniform, stable systems because they avoid scheduling overhead.

What Problem Does This Paper Address?

Communication networks have finite limits. If too many devices try to talk at once, delays skyrocket and messages get lost. To prevent this, we use scheduling.

  1. Periodic (Round-Robin): Simple and predictable. Every agent gets a slot in a circular order. There is zero overhead, meaning all available bandwidth goes to data. However, it’s "dumb"—it might give a slot to a system that is perfectly stable while ignoring one that is crashing.
  2. Event-Based (Predictive Triggering): "Smart" but costly. It identifies which systems have the highest estimation error and prioritizes them. To do this, agents must first send their "priority levels" to a scheduler. This exchange eats up bandwidth, leaving fewer slots available for actual state measurements.

The authors seek to answer: At what point does the benefit of prioritizing "needy" systems outweigh the loss of total communication slots?

Methods

The authors focus on Linear Time-Invariant (LTI) systems with process noise. They compare two specific strategies:

  • Predictive Triggering: This method predicts which systems will need to communicate soon and allocates slots accordingly. Because of the overhead of sharing priorities, it can only afford $K_{pred}$ slots per round.
  • Round-Robin: A standard periodic approach that uses the full bandwidth, providing $K_{per}$ slots (where $K_{per} > K_{pred}$).

The Core Contribution: The authors derive upper bounds for the expected quadratic estimation error for both schemes.

  • For Periodic, the error is bounded by the maximum time an agent has to wait for its turn ($T_{per}$).
  • For Predictive, the bound is more complex, factoring in the "gain" of the systems (how fast their error grows) and the reduced number of slots.

By comparing these two bounds, they create an Adaptive Scheme that can switch between periodic and event-based communication in real-time as system dynamics change.

Datasets

The paper evaluates the theory using two primary simulations:

  1. Synthetic LTI Systems: A setup with 20 randomized linear systems. This allowed the authors to precisely control noise levels ($\sigma$) to see how the trade-off shifts when one or many systems become "noisy."
  2. Simulated Cart-Poles: 20 cart-pole systems simulated using the MuJoCo physics engine. This provided a more complex, non-linear environment to test if the LTI-based theory holds up in more realistic scenarios.

Key Insights and Ideas

  • The Homogeneity Penalty: If all systems in a network are identical and behaving well, event-based communication is actually a bad choice. The overhead of scheduling provides no useful information because everyone is equally "needy," but you end up with fewer total messages.
  • Heterogeneity is the Sweet Spot: Event-based methods shine when systems are different. If five drones in a swarm are fighting a wind gust while fifteen are in calm air, predictive triggering can "steal" slots from the calm drones and give them to the ones struggling.
  • The "Aha" Moment on Bandwidth: The paper quantifies the "cost of intelligence." If the scheduling overhead reduces your available slots too much (e.g., from 5 slots down to 2), even the smartest algorithm can't compensate for the lack of raw data.

Key Results

  • Synthetic Test: In the beginning (homogeneous systems), Periodic and Predictive performed similarly. When the noise of a single system was increased, Predictive triggering significantly reduced the total network error. However, when many systems became noisy, the Periodic scheme became superior again because it could handle more total messages.
  • Cart-Pole Test: The authors simulated 20 carts and then suddenly reduced the mass of 5 of them (simulating a change in dynamics).
  • Before the change: Periodic was slightly better or equal.
  • After the change: Predictive triggering significantly outperformed Periodic, as it could prioritize the 5 carts that were now more volatile due to their lower mass.
  • Adaptive Success: The proposed adaptive scheme successfully identified these shifts and switched to the optimal strategy every time, maintaining the lowest possible error across the entire simulation.

Limitations and Future Work

  • Worst-Case vs. Average: The current criterion is based on worst-case error bounds. The authors note that an "average performance" measure might be more practical for some applications but harder to derive mathematically.
  • Simple Baselines: The periodic baseline used was a very basic round-robin. Future research could compare predictive triggering against more sophisticated periodic schemes that are pre-tuned for known system differences.
  • Network Reliability: The study assumes a reliable network where scheduled slots always result in successful delivery. In real-world wireless scenarios, packet loss is a major factor that would add another layer to the trade-off.
  • Disturbance Types: The authors suggest looking into how these methods handle large-scale disturbances (like a wind gust hitting an entire swarm) versus localized ones.

Summary updated

Abstract

Event-based methods carefully select when to transmit information to enable high-performance control and estimation over resource-constrained communication networks. However, they come at a cost. For instance, event-based communication induces a higher computational load and increases the complexity of the scheduling problem. Thus, in some cases, allocating available slots to agents periodically in circular order may be superior. In this article, we discuss, for a specific example, when the additional complexity of event-based methods is beneficial. We evaluate our analysis in a synthetical example and on 20 simulated cart-pole systems.
BibTeX
@inproceedings{baumann2023trade,
  author = {Dominik Baumann and Thomas B. Schön},
  title = {{On the trade-off between event-based and periodic state estimation under bandwidth constraints}},
  booktitle = {IFAC-PapersOnLine},
  year = {2023},
  doi = {10.1016/j.ifacol.2023.10.168},
}

Loading related content...

Figure