def step(self): """ Description: Moves the system dynamics one time-step forward. Args: None Returns: The next value in the ARMA time-series. """ assert self.initialized self.T += 1 if(self.T == 5000): self.d = 0 self.delta_i_x = None if(self.noise_list is not None): self.x, self.delta_i_x, self.noise, x_new = self._step(self.x, self.delta_i_x, self.noise, self.noise_list[self.q + self.T - 1]) else: if(self.noise_distribution == 'normal'): self.x,self.delta_i_x, self.noise, x_new = self._step(self.x, self.delta_i_x,self.noise, \ self.noise_magnitude * random.normal(generate_key(), shape=(self.n,))) elif(self.noise_distribution == 'unif'): self.x, self.delta_i_x, self.noise, x_new = self._step(self.x, self.delta_i_x,self.noise, \ self.noise_magnitude * random.uniform(generate_key(), shape=(self.n,), minval=-1., maxval=1.)) return x_new
def test_rnn(steps=100, show_plot=True):
T = steps
p, q = 3, 3
n = 1
problem = tigerforecast.problem("ARMA-v0")
y_true = problem.initialize(p=p, q=q, n=1)
method = tigerforecast.method("RNN")
method.initialize(n=1, m=1, l=3, h=1)
loss = lambda pred, true: np.sum((pred - true)**2)
results = []
for i in range(T):
u = random.normal(generate_key(), (n, ))
y_pred = method.predict(u)
y_true = problem.step()
results.append(loss(y_true, y_pred))
method.update(y_true)
if show_plot:
plt.plot(results)
plt.title("RNN method on LDS problem")
plt.show(block=False)
plt.pause(3)
plt.close()
print("test_rnn passed")
return
def step(self):
"""
Description: Moves the system dynamics one time-step forward.
Args:
u (numpy.ndarray): control input, an n-dimensional real-valued vector.
Returns:
A new observation from the LDS.
"""
assert self.initialized
self.T += 1
x = random.normal(generate_key(), shape=(self.n, ))
self.h, y = self._step(
x, self.h, (random.normal(generate_key(), shape=(self.d, )),
random.normal(generate_key(), shape=(self.m, ))))
return x, y
def initialize(self, n, m, k, T, eta, R_M):
"""
Description: Initialize the (non-existent) hidden dynamics of the method
Args:
None
Returns:
None
"""
self.initialized = True
self.n, self.m, self.k, self.T = n, m, k, T
self.eta, self.R_M = eta, R_M
self.k_prime = n * k + 2 * n + m
self.M = rand.uniform(generate_key(), shape=(self.m, self.k_prime))
if (4 * k > T):
raise Exception("Method parameter k must be less than T/4")
self.X = np.zeros((n, T))
self.Y = np.zeros((m, T))
self.X_sim = None
self.t = 0
self.y_hat = None
self.k_values, self.k_vectors = self.eigen_pairs(T, k)
self.eigen_diag = np.diag(self.k_values**0.25)
@jax.jit
def _update_x(X, x):
new_x = np.roll(X, 1, axis=1)
new_x = jax.ops.index_update(new_x, jax.ops.index[:, 0], x)
return new_x
self._update_x = _update_x
self.params = {}
def initialize(self, n, m, d, noise=1.0):
"""
Description: Randomly initialize the hidden dynamics of the system.
Args:
n (int): Input dimension.
m (int): Observation/output dimension.
d (int): Hidden state dimension.
noise (float): Default value 1.0. The magnitude of the noise (Gaussian) added
to both the hidden state and the observable output.
Returns:
The first value in the time-series
"""
self.initialized = True
self.T = 0
self.max_T = -1
self.has_regressors = True
self.n, self.m, self.d, self.noise = n, m, d, noise
# shrinks matrix M such that largest eigenvalue has magnitude k
normalize = lambda M, k: k * M / np.linalg.norm(M, ord=2)
# initialize matrix dynamics
self.A = random.normal(generate_key(), shape=(d, d))
self.B = random.normal(generate_key(), shape=(d, n))
self.C = random.normal(generate_key(), shape=(m, d))
self.D = random.normal(generate_key(), shape=(m, n))
self.h = random.normal(generate_key(), shape=(d, ))
# adjust dynamics matrix A
self.A = normalize(self.A, 1.0)
self.B = normalize(self.B, 1.0)
self.C = normalize(self.C, 1.0)
self.D = normalize(self.D, 1.0)
def _step(u, h, eps):
eps_h, eps_y = eps
next_h = np.dot(self.A, h) + np.dot(self.B, u) + self.noise * eps_h
y = np.dot(self.C, next_h) + np.dot(self.D, u) + self.noise * eps_y
return (next_h, y)
self._step = jax.jit(_step)
return self.step()
def initialize(self, params = {'T': 10000}, p=3, q=3, n = 1, d=2, noise_list = None, c=0, noise_magnitude=0.1, noise_distribution = 'normal'):
"""
Description: Randomly initialize the hidden dynamics of the system.
Args:
p (int/numpy.ndarray): Autoregressive dynamics. If type int then randomly
initializes a Gaussian length-p vector with L1-norm bounded by 1.0.
If p is a 1-dimensional numpy.ndarray then uses it as dynamics vector.
q (int/numpy.ndarray): Moving-average dynamics. If type int then randomly
initializes a Gaussian length-q vector (no bound on norm). If p is a
1-dimensional numpy.ndarray then uses it as dynamics vector.
n (int): Dimension of values.
c (float): Default value follows a normal distribution. The ARMA dynamics
follows the equation x_t = c + AR-part + MA-part + noise, and thus tends
to be centered around mean c.
Returns:
The first value in the time-series
"""
self.initialized = True
self.T = 0
self.max_T = params['T']
self.n = n
self.d = d
if type(p) == int:
phi = random.normal(generate_key(), shape=(p,))
self.phi = 0.99 * phi / np.linalg.norm(phi, ord=1)
else:
self.phi = p
if type(q) == int:
self.psi = random.normal(generate_key(), shape=(q,))
else:
self.psi = q
if(type(self.phi) is list):
self.p = self.phi[0].shape[0]
else:
self.p = self.phi.shape[0]
if(type(self.psi) is list):
self.q = self.psi[0].shape[0]
else:
self.q = self.psi.shape[0]
self.noise_magnitude, self.noise_distribution = noise_magnitude, noise_distribution
self.c = random.normal(generate_key(), shape=(self.n,)) if c == None else c
self.x = random.normal(generate_key(), shape=(self.p, self.n))
if self.d>1:
self.delta_i_x = random.normal(generate_key(), shape=(self.d-1, self.n))
else:
self.delta_i_x = None
self.noise_list = None
if(noise_list is not None):
self.noise_list = noise_list
self.noise = np.array(noise_list[0:self.q])
elif(noise_distribution == 'normal'):
self.noise = self.noise_magnitude * random.normal(generate_key(), shape=(self.q, self.n))
elif(noise_distribution == 'unif'):
self.noise = self.noise_magnitude * random.uniform(generate_key(), shape=(self.q, self.n), \
minval=-1., maxval=1.)
self.feedback=0.0
self.x_new = self.x[0]
def _step(x, delta_i_x, noise, eps):
if(type(self.phi) is list):
x_ar = np.dot(x.T, self.phi[self.T])
else:
x_ar = np.dot(x.T, self.phi)
if(type(self.psi) is list):
x_ma = np.dot(noise.T, self.psi[self.T])
else:
x_ma = np.dot(noise.T, self.psi)
if delta_i_x is not None:
x_delta_sum = np.sum(delta_i_x)
else :
x_delta_sum = 0.0
x_delta_new=self.c + x_ar + x_ma + eps
x_new = x_delta_new+x_delta_sum
next_x = np.roll(x, self.n)
next_noise = np.roll(noise, self.n)
next_x = jax.ops.index_update(next_x, 0, x_delta_new) # equivalent to self.x[0] = self.x_new
next_noise = jax.ops.index_update(next_noise, 0, eps) # equivalent to self.noise[0] = eps
next_delta_i_x=None
for i in range(d-1):
if i==0:
next_delta_i_x=jax.ops.index_update(delta_i_x, i, x_delta_new+delta_i_x[i])
else:
next_delta_i_x=jax.ops.index_update(delta_i_x, i, next_delta_i_x[i-1]+next_delta_i_x[i])
return (next_x, next_delta_i_x, next_noise, x_new)
self._step = jax.jit(_step)
if self.delta_i_x is not None:
x_delta_sum= np.sum(self.delta_i_x)
else:
x_delta_sum= 0
return self.x[0]+x_delta_sum
def test_wave_filtering(show_plot=False):
# state variables
T, n, m = 1000, 10, 10
# method variables
k, eta = 20, 0.00002 # update_steps is an optional variable recommended by Yi
R_Theta = 5
R_M = 2 * R_Theta * R_Theta * k**0.5
# generate random data (columns of Y)
hidden_state_dim = 5
h = rand.uniform(generate_key(),
minval=-1,
maxval=1,
shape=(hidden_state_dim, )) # first hidden state h_0
A = rand.normal(generate_key(), shape=(hidden_state_dim, hidden_state_dim))
B = rand.normal(generate_key(), shape=(hidden_state_dim, n))
C = rand.normal(generate_key(), shape=(m, hidden_state_dim))
D = rand.normal(generate_key(), shape=(m, n))
A = (A + A.T) / 2 # make A symmetric
A = 0.99 * A / np.linalg.norm(A, ord=2)
if (np.linalg.norm(B) > R_Theta):
B = B * (R_Theta / np.linalg.norm(B))
if (np.linalg.norm(C) > R_Theta):
C = C * (R_Theta / np.linalg.norm(C))
if (np.linalg.norm(D) > R_Theta):
D = D * (R_Theta / np.linalg.norm(D))
if (np.linalg.norm(h) > R_Theta):
h = h * (R_Theta / np.linalg.norm(h))
# input vectors are random data
X = rand.normal(generate_key(), shape=(n, T))
# generate Y according to predetermined matrices
Y = []
for t in range(T):
Y.append(
C.dot(h) + D.dot(X[:, t]) +
rand.truncated_normal(generate_key(), 0, 0.1, shape=(m, )))
h = A.dot(h) + B.dot(X[:, t]) + rand.truncated_normal(
generate_key(), 0, 0.1, shape=(hidden_state_dim, ))
Y = np.array(Y).T # list to numpy matrix
method = tigerforecast.method("WaveFiltering")
method.initialize(n, m, k, T, eta, R_M)
# loss = lambda y_true, y_pred: (y_true - y_pred)**2
loss = lambda y_true, y_pred: (np.linalg.norm(y_true - y_pred))**2
lastvalue_method = tigerforecast.method("LastValue")
lastvalue_method.initialize()
results = []
lastvalue_results = []
for i in range(T):
# print(i)
cur_y_pred = method.predict(X[:, i])
#print(method.forecast(cur_x, 3))
cur_y_true = Y[:, i]
cur_loss = loss(cur_y_true, cur_y_pred)
results.append(cur_loss)
method.update(cur_y_true)
lastvalue_cur_y_pred = lastvalue_method.predict(X[:, i])
lastvalue_cur_y_true = Y[:, i]
lastvalue_cur_loss = loss(lastvalue_cur_y_true, lastvalue_cur_y_pred)
lastvalue_results.append(lastvalue_cur_loss)
# print("test_wave_filtering passed")
print(np.linalg.norm(X[:, -1]))
print(np.linalg.norm(Y[:-1]))
print(results[-10:-1])
if show_plot:
plt.plot(results)
plt.title("WaveFiltering method on random data")
plt.show(block=True)
plt.plot(lastvalue_results)
plt.title("LastValue method on random data")
plt.show(block=True)
