Efficient Online Learning in Resource-Constrained Automation Environments

Abstract

Machine learning solutions have proven highly effective for various tasks in recent years. However, their use in an automation environment requires that they run locally with limited resources, in a setting called Edge Computing. At the same time, there is a need to facilitate continuous improvements and updates throughout the product lifecycle to ensure that systems are adaptable to evolving environments and under performance degradation of machines. For this reason, Incremental Learning models have become increasingly relevant due to their ability to process data in real-time, while also lifting the need to store all data in memory. However, efficiency in these models is often overlooked, with many implementations in Python resulting in a substantial memory footprint and slow execution, making the usage of such models in robotic controllers impractical due to the high cost of improving hardware. In this work, we implement an efficient online learning model called Mondrian Forests using the Rust language, achieving a 28-fold improvement in execution speed compared to the Python implementation. Additionally, we apply memory optimizations through spatial locality caching, further reducing execution time by 18%. Consequently, we measure performance using datasets from real-world industrial settings, analyzing the implications for automation.