def test_expand_multiple_dims():
x = np.arange(4).reshape(-1)
assert expand_multiple_dims(x, 1, 0).shape == (1, 4)
assert expand_multiple_dims(x, 0, 2).shape == (4, 1, 1)
y = np.arange(12).reshape(3, 4)
assert expand_multiple_dims(y, 3, 2).shape == (1, 1, 1, 3, 4, 1, 1)
def _update_mutual_info(self):
"""
Update mutual information using posterior distributions.
If there is no nedd to update mutual information, it ends.
The flag to update mutual information must be true only when
posteriors are updated in :code:`update(design, response)` method.
"""
# If there is no need to update mutual information, it ends.
if not self.flag_update_mutual_info:
return
# Calculate the marginal log likelihood.
lp = expand_multiple_dims(self.log_post, 1, 1)
mll = logsumexp(self.log_lik + lp, axis=1)
self.marg_log_lik = mll # shape (num_design, num_response)
# Calculate the marginal entropy and conditional entropy.
self.ent_marg = -np.sum(np.exp(mll) * mll, -1) # shape (num_designs,)
self.ent_cond = np.sum(self.post * self.ent_obs,
axis=1) # shape (num_designs,)
# Calculate the mutual information.
self.mutual_info = self.ent_marg - \
self.ent_cond # shape (num_designs,)
# Flag that there is no need to update mutual information again.
self.flag_update_mutual_info = False
def reset(self):
"""
Reset the engine as in the initial state.
"""
self.y_obs = np.array(self.task.responses)
self.p_obs = self._compute_p_obs()
self.log_lik = ll = self._compute_log_lik()
lp = np.ones(self.grid_param.shape[0])
self.log_prior = lp - logsumexp(lp)
self.log_post = self.log_prior.copy()
lp = expand_multiple_dims(self.log_post, 1, 1)
mll = logsumexp(self.log_lik + lp, axis=1)
self.marg_log_lik = mll # shape (num_design, num_response)
self.ent_obs = -np.multiply(np.exp(ll), ll).sum(-1)
self.ent_marg = None
self.ent_cond = None
self.mutual_info = None
self.eligibility_trace = np.zeros(self.grid_design.shape[0])
self.flag_update_mutual_info = True