def eps_greedy_choice_linear(self, state, epsilon):
Qa = np.zeros(2)
# epsilon greedy policy
if random.random() > epsilon:
for action in Actions.get_values():
phi = self.feature_computation(state, action)
Qa[action] = sum(phi*self.theta)
a_next = Actions.get_action(np.argmax(Qa))
else:
a_next = Actions.hit if random.random()<0.5 else Actions.stick
phi = self.feature_computation(state, a_next)
my_Qa = sum(phi*self.theta)
return [a_next, my_Qa]
def eps_greedy_choice_linear(self, state, epsilon):
Qa = np.zeros(2)
# epsilon greedy policy
if random.random() > epsilon:
for action in Actions.get_values():
phi = self.feature_computation(state, action)
Qa[action] = sum(phi * self.theta)
a_next = Actions.get_action(np.argmax(Qa))
else:
a_next = Actions.hit if random.random() < 0.5 else Actions.stick
phi = self.feature_computation(state, a_next)
my_Qa = sum(phi * self.theta)
return [a_next, my_Qa]
def eps_greedy_choice(self, state):
# collect visits
try:
visits_to_state = sum(self.N[state.dl_sum-1, state.pl_sum-1, :])
except:
visits_to_state = 0
# compute epsilon
curr_epsilon = self.n0 / (self.n0 + visits_to_state)
# epsilon greedy policy
if random.random() < curr_epsilon:
return Actions.hit if random.random()<0.5 else Actions.stick
else:
return Actions.get_action(np.argmax(self.Q[state.dl_sum-1, state.pl_sum-1, :]))
def eps_greedy_choice(self, state):
# collect visits
try:
visits_to_state = sum(self.N[state.dl_sum - 1,
state.pl_sum - 1, :])
except:
visits_to_state = 0
# compute epsilon
curr_epsilon = self.n0 / (self.n0 + visits_to_state)
# epsilon greedy policy
if random.random() < curr_epsilon:
return Actions.hit if random.random() < 0.5 else Actions.stick
else:
return Actions.get_action(
np.argmax(self.Q[state.dl_sum - 1, state.pl_sum - 1, :]))