def __call__(self, action=None):
if action == None:
self.curr_pos = self.start_pos
self.episode_steps = 0
self.start_episode()
return self.state()
else:
self.episode_steps += 1
assert action in self.actions
r, c = self.curr_pos
p = self.correct_action_probability
N = len(self.actions)
distr = array([(1 - p) / (N - 1)] * N)
distr[self.actions.index(action)] = p
a = utils.weighted_sample(distr)
dr, dc = self.action_map[self.actions[a]]
if self.move_okay(r + dr, c + dc):
r, c = self.curr_pos = (r + dr, c + dc)
if (r, c) == self.goal_pos:
self.verbose("!!! GOAL !!!")
return rl.TERMINAL_STATE, self.goal_reward
elif self.timeout and self.episode_steps > self.timeout:
return rl.TERMINAL_STATE, self.step_reward
else:
return self.state(), self.step_reward
def policy(self,sensation):
"""
Given a sensation, return an action. Uses
self.action_selection to get a distribution over the agent's
actions. Uses self.applicable_actions to prevent selecting
inapplicable actions.
Returns 0 if is_terminal(sensation).
"""
if not is_terminal(sensation):
actions = self.applicable_actions(sensation)
return actions[weighted_sample(self.policy_fn(sensation,actions))]
else:
# In the terminal state, the action is irrelevant
return 0
def policy(self, sensation):
"""
Given a sensation, return an action. Uses
self.action_selection to get a distribution over the agent's
actions. Uses self.applicable_actions to prevent selecting
inapplicable actions.
Returns 0 if is_terminal(sensation).
"""
if not is_terminal(sensation):
actions = self.applicable_actions(sensation)
return actions[weighted_sample(self.policy_fn(sensation, actions))]
else:
# In the terminal state, the action is irrelevant
return 0