rapidsai / cuml

cuML - RAPIDS Machine Learning Library

https://docs.rapids.ai/api/cuml/stable/

Apache License 2.0

4.24k stars 532 forks source link

[ENH] Seasonal ARIMA: remove complexity bottleneck (initialization of the Kalman filter) #1696

Closed Nyrio closed 4 years ago

Nyrio commented 4 years ago

The performance of seasonal ARIMA in PR #1586 is limited by the initialization of the Kalman filter (in particular the matrix P). Solving the Lyapunov/Sylvester equation is very expensive.

We can speed it up with the Bartels-Stewart algorithm if we manage to make a good implementation of the batched real Schur decomposition. Iterative methods to approximate the solution are also a possibility. More info on: https://en.wikipedia.org/wiki/Bartels%E2%80%93Stewart_algorithm

Nyrio commented 4 years ago

1788

Nyrio commented 4 years ago

Solved, closing