UC Irvine

Decision-making and real-time prediction in non-stationary and dynamic environments present significant challenges for the application of machine learning and artificial intelligence in operational settings. In these environments, deploying pre-trained models for real-time predictions on data streams often yields suboptimal results due to potential distribution shifts between the training data and the incoming data stream. Moreover, the computational complexity of online decision-making algorithms is crucial, as timely decisions must be made before new data points are received. Furthermore, these algorithms should be designed to maintain affordable memory complexity, as the sheer volume of streaming data cannot be stored in batch. This thesis presents several novel online decision-making and real-time prediction algorithms that effectively address the challenges of non-stationary environments while ensuring computational and memory efficiency.

Novel online multi-kernel learning algorithms for real-time prediction on data streams are introduced, with theoretical guarantees and experimental results demonstrating their advantages in both accuracy and computational efficiency over state-of-the-art online kernel learning algorithms. In online decision-making, feedback is typically observed after each decision round through interactions with the environment. This thesis introduces novel algorithms for online decision-making under uncertain observations. Theoretical analysis and experimental results demonstrate the effectiveness of the proposed algorithms in handling uncertainty during decision-making.

Meanwhile, federated learning is well-regarded for its ability to facilitate distributed model training while ensuring data privacy, as clients retain their data without sharing it with the central server that coordinates the collaboration. Most previous research assumes that clients have static batches of training data. However, in many cases, clients need to make real-time predictions on streaming data in non-stationary environments. Another significant challenge in federated learning is the heterogeneity of data distribution among clients. To address these challenges, this thesis introduces novel personalized online federated learning algorithms. The proposed algorithms enjoy theoretical guarantees, and experiments on real datasets demonstrate their superiority over existing online federated learning approaches in real-time prediction, particularly in the presence of heterogeneous data distribution among clients.