def __init__(self, rng: np.random.RandomState, arms: List[Arm],
n_jobs: int, backend: Optional[str],
lp: Union[_EpsilonGreedy, _Linear, _Random, _Softmax,
_ThompsonSampling,
_UCB1], n_clusters: Num, is_minibatch: bool):
super().__init__(rng, arms, n_jobs, backend)
self.n_clusters = n_clusters
if is_minibatch:
self.kmeans = MiniBatchKMeans(n_clusters,
random_state=rng.get_state()[1][0])
else:
self.kmeans = KMeans(n_clusters,
random_state=rng.get_state()[1][0])
# Create the list of learning policies for each cluster
# Deep copy all parameters of the lp objects, except refer to the originals of rng and arms
self.lp_list = [deepcopy(lp) for _ in range(self.n_clusters)]
for c in range(self.n_clusters):
self.lp_list[c].rng = rng
self.lp_list[c].arms = arms
self.decisions = None
self.rewards = None
self.contexts = None
# Initialize the arm expectations to nan
# When there are neighbors, expectations of the underlying learning policy is used
# When there are no neighbors, return nan expectations
reset(self.arm_to_expectation, np.nan)
def temp_seed(rng: np.random.RandomState,
seed: Optional[Union[int, Tuple[int, ...]]]):
"""A context manager for temporarily adjusting the random seed."""
if seed is None:
try:
yield
finally:
pass
else:
state = rng.get_state()
rng.seed(seed)
try:
yield
finally:
rng.set_state(state)
def serialize_random_state(random_state: np.random.RandomState) -> Tuple[int, List, int, int, int]:
(rnd0, rnd1, rnd2, rnd3, rnd4) = random_state.get_state()
rnd1 = rnd1.tolist()
return rnd0, rnd1, rnd2, rnd3, rnd4
