def test_calc_belief_matrix():
sm = np.array([[.3, .6], [.7, .4]])
wp = np.array([.5, .5])
beliefs = fwd.calc_belief_matrix(sm, wp)
correct_beliefs = np.array([[0.33333333, 0.63636364],
[ 0.66666667, 0.36363636]])
npt.assert_allclose(beliefs, correct_beliefs, rtol=1e-04)
def prob_meta_signals(given, sm, wp, noise, params, belief_matrix=None,
meta_matrix=None):
if not params.use_metas:
return np.ones((sm.shape[0], 1))
if params.prior_wp_sm in ["none", "signal_assumption",
"gen_bij_infer_all", "bijection"]:
if belief_matrix is None:
belief_matrix = fwd.calc_belief_matrix(sm, wp)
if meta_matrix is None:
meta_matrix = fwd.calc_meta_matrix(sm, wp, belief_matrix)
if (params.use_expertise and params.ppl_know_expertise):
bayes_s0_sig = np.array([unanimity_metas(bm, mm, params.prior_wp_sm)
for bm, mm in zip(belief_matrix, meta_matrix)])
bayes_s0_sig0 = bayes_s0_sig[:,0]
bayes_s0_sig1 = bayes_s0_sig[:,1]
else:
bayes_s0_sig = unanimity_metas(belief_matrix, meta_matrix,
params.prior_wp_sm)
bayes_s0_sig0 = bayes_s0_sig[0]
bayes_s0_sig1 = bayes_s0_sig[1]
pas = np.zeros((2, given.shape[1]))
pas[0,:] = prob_noisy_vec(given[0,:], bayes_s0_sig0,
params.meta_noise_type, noise)
#if bayes_s0_sig[1] == bayes_s0_sig[0]:
# pas[1,:] = pas[0,:]
#else:
pas[1,:] = prob_noisy_vec(given[0,:], bayes_s0_sig1,
params.meta_noise_type, noise)
#pas = mask_nan_probs(pas, given)
return pas
def reports_ll_person(beliefs, meta, world, sm, wp, own_noise, meta_noise,
expertise, bias, params, bmat=None, mmat=None, debug=None):
if params.use_expertise:
if params.ppl_know_expertise:
if params.orig_expertise_code:
bmat = [fwd.calc_belief_matrix_expertise(sm, wp, ee)
for ee in expertise]
else:
bmat = [fwd.calc_belief_matrix_expertise_min(sm, wp, ee, bias)
for ee in expertise]
mmat = [fwd.calc_meta_matrix(sm, wp, bm) for bm in bmat]
if bmat is None:
bmat = fwd.calc_belief_matrix(sm, wp)
if mmat is None:
mmat = fwd.calc_meta_matrix(sm, wp, bmat)
prob_signals = calc_prob_signals(sm, world, expertise, bias, params)
pbs_vec = prob_beliefs_signals(beliefs, sm, wp, own_noise, params, bmat)
pms_vec = prob_meta_signals(meta, sm, wp, meta_noise, params, bmat, mmat)
ll = np.log((prob_signals * pbs_vec * pms_vec).sum(axis=0))
if debug is not None:
1/0
#check_log = np.log(prob_signals) + np.log(pbs_vec) + np.log(pms_vec)
#check_exp = np.exp(check_log)
#check_final = np.log(np.nansum(check_exp, 0))
return ll
def bijection(self, sm, wp):
beliefs = fwd.calc_belief_matrix(sm, wp)
bijection = True
for probs in beliefs.T:
if sum(np.max(probs) == probs) > 1:
bijection = False
break
argmaxs = np.argmax(beliefs, axis=0)
if len(np.unique(argmaxs)) < sm.shape[0]:
bijection = False
return bijection, argmaxs
def prob_beliefs_signals(given, sm, wp, noise, params, belief_matrix=None):
if not params.use_beliefs:
return np.ones((sm.shape[0], 1))
if belief_matrix is None:
belief_matrix = fwd.calc_belief_matrix(sm, wp)
if params.binary_beliefs:
pas = prob_binary_beliefs_signals(given, noise, params, belief_matrix)
else:
if params.votes_confs:
pas = prob_votes_confs(given, noise, params, belief_matrix)
else:
pas = prob_non_binary_beliefs(given, noise, params, belief_matrix)
#pas = mask_nan_probs(pas, given)
return pas