News

Title: Optimizing Hierarchical Classification Problems for Resource-Constrained Systems Using Reinforcement Learning Methods

on resource-constrained embedded devices. Approximating machine learning models

to deploy them at the edge might compromise the predictive accuracy. On the other

hand, deploying complex machine learning models at the edge might cause hardware

overheating and decrease the life span of the edge devices. In [1 ], the authors propose

hierarchical classification to achieve a trade-off between predictive accuracy and energy consumption at inference. The authors design a reinforcement learning framework to select the optimal classifier at every level of the hierarchical classification problem. However, the authors only consider the estimation of the energy. We introduce the

HierarchicalEdge framework that extends the previous method by considering real

energy measurements. In addition, the HierarchicalEdge framework implements feature selection, as well as model selection. The classifiers selected by the HierarchicalEdge framework consume low energy at inference while maintaining high predictive accuracy.

This thesis shows that the HierarchicalEdge framework is superior to both flat classifiers

and the hierarchical classifier selected by the reinforcement learning framework of [1 ].

Additionally, the thesis shows that the classifiers and feature sets selected by the

HierarchicalEdge framework showed better accuracy and energy consumption trade-off

Sherwood Polter, and Qing Dong. Hierarchical fault classification for resource

constrained systems. Mechanical Systems and Signal Processing, 134:106–266,