Linear dynamical systems (LDSs) are a class of time-series models widely used in robotics, finance, engineering, and meteorology. In it's general form (when state transition dynamics are unknown), learning LDS is a classic non-convex problem, typically tackled with heuristics like gradient descent ("backpropagation through time") or the EM algorithm. I will present our new "spectral filtering" approach to the identification and control of discrete-time general LDSs with multi-dimensional inputs, outputs, and a latent state. This approach yields a simple, efficient, and practical algorithm for low-regret prediction (i.e. asymptotically vanishing MSE).
The talk will cover a series of results, which are joint work with Elad Hazan, Cyril Zhang, Sanjeev Arora, Holden Lee, and Yi Zhang.