View source on GitHub
|
Updates the inverse using the Woodbury matrix identity.
tf_agents.bandits.policies.linalg.update_inverse(
a_inv: tf_agents.typing.types.Float,
x: tf_agents.typing.types.Float
) -> tf_agents.typing.types.Float
Given a matrix A of size d-by-d and a matrix X of size k-by-d, this
function computes the inverse of B = A + X^T X, assuming that the inverse of
A is available.
Reference:
https://en.wikipedia.org/wiki/Woodbury_matrix_identity
Args | |
|---|---|
a_inv
|
a Tensor of shape [d, d]. This is the current inverse of A.
|
x
|
a Tensor of shape [k, d].
|
Returns | |
|---|---|
The update that needs to be added to 'a_inv' to compute the inverse.
If x is empty, a zero matrix is returned.
|