def __init__(self, prior: np.ndarray= None):
"""
@ Args
prior: a (n, 2) ndarray with beta priors. prior[:,0] is the first concentration
If given, `n` has to be the same as number of valid actions
"""
prior = torch.tensor(prior, dtype= torch.float32)
save__init__args(locals())
def __init__(self,
win_probs: List[float],
):
"""
@ Args:
win_probs: telling the winning probablilty of each arm, which also gives the number of arms
"""
win_probs = np.array(win_probs)
save__init__args(locals(), underscore=True)
self._action_space = IntBox(0, len(self._win_probs))
self._observation_space = IntBox(0, 1) # This serves no purpose, just to meet the interface
self.BanditEnvInfo = namedtuple("BanditEnvInfo",
[*BanditEnvInfoBase._fields] + ["arm{}".format(i) for i in range(len(win_probs))])
def __init__(
self,
EnvCls,
env_kwargs,
traj_len: int = 1,
batch_size: int = 1,
**kwargs,
):
"""
@ Args:
traj_len / T: maximum transitions in one sample trajectory.
If under some circumstance not reached, reset of them will be filled with zeros.
batch_size / B: the batch size of one sample operation.
"""
save__init__args(locals())
def __init__(
self,
algo: AlgoBase,
agent: AgentBase,
sampler: SamplerBase,
affinity,
max_train_epochs: int,
log_interval: int = 1,
**kwargs,
):
"""
@ Args
max_train_epochs: The maximum number of training epoches,
One epoch is one-time of calling algo.train()
log_interval: The interval of actually logging into file
affinity: Incase you run multiple experiment on one machine
"""
self.algo = algo
self.agent = agent
self.sampler = sampler
save__init__args(locals())
def __init__(self, c= 1.0):
""" As UCB algorithm described, you need to provide a factor c
"""
save__init__args(locals())
def __init__(self, epsilon):
save__init__args(locals())
def __init__(self,
random_init= False,
beta= 1.0, # coefficient for likelyhood
b: float= None, # if None, constantly update baseline; or keep baseline constant
):
save__init__args(locals())