AdverserialPolicy¶
- class itergp.methods.policies.AdverserialPolicy(base_policy)¶
Bases:
LinearSolverPolicyAdverserial policy.
Policy returning actions which are orthogonal to the data causing no update to the posterior mean except for the last iteration. The actions are given by \(s_i = (I - y(y^\top y)^{-1}y^\top)\tilde{s}_i\), where \(\tilde{s}_i\) are arbitrary linearly independent vectors.
- Parameters
base_policy (policies.LinearSolverPolicy) – Policy which generates \(\tilde{s}_i\) which then are transformed to be orthogonal to \(y\).
- Return type
None
Attributes Summary
Methods Summary
__call__(solver_state, rng)Return an action for a given solver state.
Attributes Documentation
- base_policy¶
Methods Documentation
- __call__(solver_state, rng)[source]¶
Return an action for a given solver state.
- Parameters
solver_state (probnum.linalg.solvers.LinearSolverState) – Current state of the linear solver.
rng (Optional[SeedSequence]) – Random number generator.
- Returns
Next action to take.
- Return type
action