A Novel Resource-aware Tensor Decomposition Design Based on Reinforcement Learning

Abstract

Tensor decomposition is a promising solution for analyzing and classifying multidimensional data streams in long-term monitoring of physical systems. However, the key challenge is to continuously perform the expensive tensor decomposition to ensure the effective classification of the time-evolving data as the system evolves over time. This paper explores a novel resource-aware tensor decomposition framework using reinforcement learning (RL). The proposed framework is developed to update a tensor decomposition-based classifier only using the data that is important to the evolving nature of the system. It learns an RL agent to look-ahead and identifies the data that is important to the classifier update and performs the tensor decomposition on those data streams. The proposed resource-aware framework was evaluated using synthetic data and motion data from patients with Parkinson's disease and indicated a significant performance in both classification accuracy and number of tensor decompositions compared to a continuous update approach.