Continuous Model Adaptation Using Online Meta-learning

The rapid development of the internet of things and computing technologies makes it possible for researchers to have access to big data and develop data-driven models for complex engineering system monitoring and control in a variety of areas, while few works have been done investigating how to intelligently control hyperparameter settings through continuously updating the model with real-time data input. In our work, a novel online meta-learning framework is proposed for real-time data input pattern shifts’ digestion and provide accurate prediction and classification results compared to existing base-learning and meta-learning algorithms; A novel DDPG-based framework is developed later to combine model selection and hyperparameter optimization, formulate them into a sequential Markov decision making problem, which quickly indicates model suggested to be used for data sequence input and adaptively controls model hyperparameters with real-time response and provide better outcomes.