def test_fpi_inference(self):
num_obs = [2, 4]
num_states = [2, 2]
num_control = [2, 2]
A = utils.random_A_matrix(num_obs, num_states)
B = utils.random_B_matrix(num_states, num_control)
C = utils.obj_array_zeros([num_ob for num_ob in num_obs])
C[1][0] = 1.0
C[1][1] = -2.0
agent = Agent(A=A, B=B, C=C, control_fac_idx=[1])
o, s = [0, 2], [0, 0]
qx = agent.infer_states(o)
agent.infer_policies()
action = agent.sample_action()
self.assertEqual(len(action), len(num_control))
def test_mmp_active_inference(self):
"""
Tests to make sure whole active inference loop works (with various past and future
inference/policy horizons).
@TODO: Need to check this against a MATLAB output, where
the sequence of all observations / actions / generative model
parameters are used (with deterministic action sampling and
pre-determined generative process outputs - i.e. no effects of action)
"""
num_obs = [3, 2]
num_states = [4, 3]
num_control = [1, 3]
A = utils.random_A_matrix(num_obs, num_states)
B = utils.random_B_matrix(num_states, num_control)
C = utils.obj_array_zeros(num_obs)
C[1][0] = 1.0
C[1][1] = -2.0
agent = Agent(A=A,
B=B,
C=C,
control_fac_idx=[1],
inference_algo="MMP",
policy_len=2,
inference_horizon=3)
T = 10
for t in range(T):
o = [
np.random.randint(num_ob) for num_ob in num_obs
] # just randomly generate observations at each timestep, no generative process
qx = agent.infer_states(o)
agent.infer_policies()
action = agent.sample_action()
print(agent.prev_actions)
print(agent.prev_obs)
def test_mmp_inference(self):
num_obs = [2, 4]
num_states = [2, 2]
num_control = [2, 2]
A = utils.random_A_matrix(num_obs, num_states)
B = utils.random_B_matrix(num_states, num_control)
C = utils.obj_array_zeros(num_obs)
C[1][0] = 1.0
C[1][1] = -2.0
agent = Agent(A=A,
B=B,
C=C,
control_fac_idx=[1],
inference_algo="MMP",
policy_len=5,
inference_horizon=1)
o = [0, 2]
qx = agent.infer_states(o)
print(qx[0].shape)
print(qx[1].shape)
def test_update_posterior_states_v2(self):
"""
Testing our SPM-ified version of `run_MMP` with
1 hidden state factor & 1 outcome modality, at a random fixed
timestep during the generative process
"""
past_len = 3
future_len = 4
num_policies = 5
num_states = [6, 7, 8]
num_controls = [9, 10, 11]
num_obs = [12, 13, 14]
num_modalities = len(num_obs)
A = random_A_matrix(num_obs, num_states)
B = random_B_matrix(num_states, num_controls)
prev_obs = [rand_onehot_obs(num_obs) for _ in range(past_len)]
prev_actions = np.array([rand_controls(num_controls) for _ in range(past_len)])
policies = [
np.array([rand_controls(num_controls) for _ in range(future_len)])
for _ in range(num_policies)
]
prior = rand_dist_states(num_states)
qs_seq_pi, VFE_policies = update_posterior_states_v2(A, B, prev_obs, policies, prev_actions, prior=prior, policy_sep_prior = False)
qs_seq_pi_future = utils.obj_array(num_policies)
for p_idx in range(num_policies):
qs_seq_pi_future[p_idx] = qs_seq_pi[p_idx][(1 + past_len) :]
# create C matrix
# horizon = len(qs_seq_pi_future[0])
# C = utils.obj_array(horizon)
# for t in range(horizon):
# C[t] = utils.obj_array(num_modalities)
# for g in range(num_modalities):
# C[t][g] = np.ones(num_obs[g])
C = utils.obj_array_uniform(num_obs)
q_pi, efe = update_posterior_policies_mmp(
qs_seq_pi_future,
A,
B,
C,
policies,
use_utility=True,
use_states_info_gain=True,
use_param_info_gain=False,
prior = None,
pA=None,
pB=None,
F = VFE_policies,
E = None,
gamma=16.0,
return_numpy=True,
)
qs_pi_curr_t = utils.obj_array(num_policies)
for p_idx in range(num_policies):
qs_pi_curr_t[p_idx] = qs_seq_pi[p_idx][past_len]
qs_bma = average_states_over_policies(qs_pi_curr_t, q_pi) # Bayesian model average of hidden states across policies
return obs
def rand_controls(num_controls):
if type(num_controls) is int:
num_controls = [num_controls]
controls = np.zeros(len(num_controls))
for i in range(len(num_controls)):
controls[i] = np.random.randint(num_controls[i])
return controls
if __name__ == "__main__":
past_len = 4
future_len = 4
num_states = [8, 12, 13]
num_controls = [12, 3, 16]
num_obs = [12, 14, 6]
A = random_A_matrix(num_obs, num_states)
B = random_B_matrix(num_states, num_controls)
prev_obs = [rand_onehot_obs(num_obs) for _ in range(past_len)]
prev_actions = np.array(
[rand_controls(num_controls) for _ in range(past_len)])
policy = np.array([rand_controls(num_controls) for _ in range(future_len)])
ll_seq = get_joint_likelihood_seq(A, prev_obs, num_states)
qs_seq = run_mmp_v2(A, B, ll_seq, policy, grad_descent=True)
for t, qs in enumerate(qs_seq):
print(f"Step {t} shape {[el.shape for el in qs]}")