Probabilistic Error Reasoning on IoT Edge Devices

Abstract

Existing IoT applications are increasingly using sensors to collect real-world measurements to make decisions. Such measurements are inherently limited by the accuracy of ADC devices, hence, introduce noise and errors. However, application developers often choose scalar data to represent sensor readings without regard to the errors associated with such data. This gives the illusion that the measurements are error-free, leading to error accumulation and false positive results. In this paper, we present a new type of programming abstraction for modeling errors and performing inference tasks in measurements of the physical world on resource-constrained IoT devices, which we call approximation variables (approxes). Using approxes does not require any changes to the programming language itself. Instead, it is designed as a suite of library functions that can be integrated directly into existing programming practices. We demonstrate how to use it in C programs. This framework makes decisions about the distributions of parameter values and inherently supports sampling and hypothesis testing to evaluate the accuracy of computational results. We compare its use to traditional programming practices and show how the library can be used to reveal uncertainty to the user, so that it can handle errors, reduce false positive results, and lead to better decision-making. These benefits make approxes a compelling and promising solution for programming with noisy sensor measurements for modern IoT applications.