Lifelong learning for multi-task learning

The document summarizes lifelong multi-task reinforcement learning algorithms. It first introduces lifelong learning and multi-task learning. It then summarizes the Efficient Lifelong Learning Algorithm (ELLA), which learns tasks sequentially by optimizing a shared latent feature matrix and task-specific weight vectors. It describes how ELLA resolves inefficiencies in this process. It also summarizes an extension of ELLA to policy gradient methods called PG-ELLA, which applies these ideas to reinforcement learning. The document concludes by noting limitations like the linear model and discussing future directions like learning the dimensionality of the latent space and using deep neural networks.