def test_equally_probable(self):
"""Test generation of equally probable values."""
# no overrides
alp = ['A', 'B', 'C', 'D']
probs = equally_probable(alp)
self.assertEqual(len(alp), len(probs))
for prob in probs:
self.assertEqual(0.25, prob)
# test with override
alp = ['A', 'B', 'C', 'D']
probs = equally_probable(alp, {'A': 0.4})
self.assertEqual(len(alp), len(probs))
self.assertAlmostEqual(1.0, sum(probs))
self.assertEqual(probs[0], 0.4)
self.assertAlmostEqual(probs[1], 0.2)
self.assertAlmostEqual(probs[2], 0.2)
self.assertAlmostEqual(probs[3], 0.2)
# test with 0.0 override
alp = ['A', 'B', 'C', 'D', 'E']
probs = equally_probable(alp, {'E': 0.0})
self.assertEqual(len(alp), len(probs))
self.assertAlmostEqual(1.0, sum(probs))
self.assertEqual(probs[0], 0.25)
self.assertAlmostEqual(probs[1], 0.25)
self.assertAlmostEqual(probs[2], 0.25)
self.assertAlmostEqual(probs[3], 0.25)
self.assertAlmostEqual(probs[4], 0.0)
# test with multiple overrides
alp = ['A', 'B', 'C', 'D']
probs = equally_probable(alp, {'B': 0.2, 'D': 0.3})
self.assertEqual(len(alp), len(probs))
self.assertAlmostEqual(1.0, sum(probs))
self.assertEqual(probs[0], 0.25)
self.assertAlmostEqual(probs[1], 0.2)
self.assertAlmostEqual(probs[2], 0.25)
self.assertAlmostEqual(probs[3], 0.3)
# test with override that's not in the alphabet
alp = ['A', 'B', 'C', 'D']
probs = equally_probable(alp, {'F': 0.4})
self.assertEqual(len(alp), len(probs))
self.assertAlmostEqual(1.0, sum(probs))
for prob in probs:
self.assertEqual(0.25, prob)
def initialize_epoch(self):
"""If a decision is made initializes the next epoch."""
try:
# First, reset the history for this new epoch
self.conjugator.reset_history()
# If there is no language model specified, mock the LM prior
# TODO: is the probability domain correct? ERP evidence is in
# the log domain; LM by default returns negative log domain.
if not self.lmodel:
# mock probabilities to be equally likely for all letters.
overrides = {BACKSPACE_CHAR: self.backspace_prob}
prior = equally_probable(self.alp, overrides)
# Else, let's query the lmodel for priors
else:
# Get the displayed state
# TODO: for oclm this should be a list of (sym, prob)
update = self.decision_maker.displayed_state
# update the lmodel and get back the priors
lm_prior = self.lmodel.state_update(update)
# normalize to probability domain if needed
if getattr(self.lmodel, 'normalized', False):
lm_letter_prior = lm_prior['letter']
else:
lm_letter_prior = norm_domain(lm_prior['letter'])
if BACKSPACE_CHAR in self.alp:
# Append backspace if missing.
sym = (BACKSPACE_CHAR, self.backspace_prob)
lm_letter_prior = sym_appended(lm_letter_prior, sym)
# convert to format needed for evidence fusion;
# probability value only in alphabet order.
# TODO: ensure that probabilities still add to 1.0
prior = [
prior_prob for alp_letter in self.alp
for prior_sym, prior_prob in lm_letter_prior
if alp_letter == prior_sym
]
# Try fusing the lmodel evidence
try:
prob_dist = self.conjugator.update_and_fuse(
{'LM': np.array(prior)})
except Exception as lm_exception:
print("Error updating language model!")
raise lm_exception
# Get decision maker to give us back some decisions and stimuli
is_accepted, sti = self.decision_maker.decide(prob_dist)
except Exception as init_exception:
print("Error in initialize_epoch: %s" % (init_exception))
raise init_exception
return is_accepted, sti