We have proposed in this project the first time series classification technique that considers accuracy, earliness, and varied lengths simultaneously. The proposed technique is fit for the problem of early classification of cardiovascular diseases based on ECG signals, containing a novel deep reinforcement learning framework, Snippet Policy Network (SPN), and a new multi-objective optimization neural network algorithm, Knee-guided Neuroevolution Algorithm (KGNA). 

We have proposed a novel deep reinforcement learning framework, Snippet Policy Network, to solve this problem. Existing studies did not well explore the solution to the classification problem of varied-length ECG signals. We are the first to define clearly and then solve this problem. Since there are two objectives that need to be optimized in the early time series classification problem, we have proposed a new multi-objective optimization neural network algorithm, the Knee-Guided Neuroevolution Algorithm. The method of the deep reinforcement learning framework trains a multi-objective optimization agent that can reasonably trade off the classification accuracy and earliness of the model, thereby enhancing the accuracy and earliness of SPN. In addition to the data mentioned above in the medical field, several sensing datasets were also used to conduct a series of experiments to verify the effectiveness and generalization of the proposed methods. The experimental results were shown to deliver the best performance in this area, holding the leading position worldwide.