class Agent(agents.random.RandomAgent(epochs)):
def __init__(self, *args):
super().__init__(*args)
self.model = SparseGaussianProcess(encoder=self.encode,
dim=dim,
shape=self.shape,
beta=beta,
M=M)
def act(self, seqs):
choices = []
t = 1 + len(self.seen) // self.batch
D = len(seqs) + len(self.seen)
beta = lambda t: 2 * np.log(D * t**2 * np.pi**2 / 3)
mu, sigma = self.model.interpolate(seqs)
seqs = np.array(seqs)
ucb = mu + 2 * np.sqrt(beta(t + 1)) * sigma
selected = np.argsort(ucb)[-int(k * self.batch):]
if k != 1.:
idx = utils.mcmc.mcmc(self.batch,
self.model.embed(seqs[selected]),
iters=1000)
selected = selected[idx]
return seqs[selected]
def observe(self, data):
super().observe(data)
self.model.fit(*zip(*self.seen.items()), epochs=epochs)
def __init__(self, *args):
super().__init__(*args)
self.model = SparseGaussianProcess(encoder=self.encode,
dim=dim,
shape=self.shape,
beta=beta,
M=M)
def __init__(self, *args):
super().__init__(*args)
self.model = SparseGaussianProcess(encoder=self.encode,
dim=dim,
shape=self.shape,
beta=beta,
M=M)
if len(self.prior):
self.model.embed.fit(*zip(*self.prior.items()),
epochs=initial_epochs)