def act(self, Q: Agent, task: Task, state):
# set the number of actions of the current task, if not set
if self.valid_actions == 0:
self.valid_actions = task.valid_actions()
# get the distribution over actions for the current state
pref = self.preferences[state]
# sample an action from the preference distribution
action = np.random.choice(self.valid_actions, 1, p=pref)
# get the greedy action according to Q
greedy = Q.max_action(state)
# update the preference distribution
pref *= (1.0 - self.beta)
pref[greedy] /= (1.0 - self.beta)
pref[greedy] += self.beta * (1.0 - pref[greedy])
return action
def epsilon_greedy(self, Q: Agent, task: Task, state, epsilon):
if np.random.rand() <= epsilon:
return random.randrange(task.valid_actions())
else:
return Q.max_action(state)