def reset(self):
obs = self.env.reset()
self.estimator.reset()
bel, _ = self._update_belief(action=None, obs=obs)
obs = to_one_hot(obs, self.n)
self.last_obs = (obs, bel)
return {'obs': obs, 'zbel': bel}
def step(self, action):
obs, reward, done, info = super(ExplicitBayesRockSample,
self).step(action)
obs['obs'] = to_one_hot(obs['obs'], self.env.nS)
info['expert'] = int(''.join(str(int(x)) for x in self.env.state[2:]),
2)
return obs, reward, done, info
def step(self, action):
obs, reward, done, info = self.env.step(action)
bel, info = self._update_belief(action, obs, **info)
obs = to_one_hot(obs, self.n)
self.last_obs = (obs, bel)
true_prob = self.env.slip_prob[Action.A]
# print(true_prob)
info['expert'] = np.where(
np.array([0.1, 0.2, 0.3, 0.4, 0.5]) == true_prob)[0][0]
return {'obs': obs, 'zbel': bel}, reward, done, info
def _augment_observation(self, action, obs, **kwargs):
# Estimate
self.estimator.estimate(action, obs, **kwargs)
mle = self.estimator.get_mle()
if not isinstance(mle, np.ndarray):
mle = np.array([mle])
if isinstance(self.env.observation_space, Discrete):
obs = to_one_hot(obs, self.env.observation_space.n)
if not isinstance(obs, np.ndarray):
obs = np.array([obs])
kwargs['mle'] = mle
if self.augment_to_obs:
return np.concatenate([obs, mle], axis=0), kwargs
else:
return obs, kwargs
def _augment_observation(self, action, obs, **kwargs):
# Estimate
self.estimator.estimate(action, obs, **kwargs)
belief = self.estimator.get_belief()
if not isinstance(belief, np.ndarray):
belief = np.array([belief])
if isinstance(self.env.observation_space, Discrete):
obs = to_one_hot(obs, self.env.observation_space.n)
if not isinstance(obs, np.ndarray):
obs = np.array([obs])
kwargs['belief'] = belief
if self.augment_to_obs:
return np.concatenate([obs, belief], axis=0), kwargs
else:
return obs, kwargs