def __init__(self):
self.card_min = 1 # min absolute val of card
self.card_max = 10 # max absolute val of card
self.dl_values = 10 # possible values for dl in state
self.pl_values = 21 # possible values for pl in state
self.act_values = len(
Actions.get_values()) # number of possible actions
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 __init__(self):
self.card_min = 1 # min absolute val of card
self.card_max = 10 # max absolute val of card
self.dl_values = 10 # possible values for dl in state
self.pl_values = 21 # possible values for pl in state
self.act_values = len(Actions.get_values()) # number of possible actions
def TD_control_linear(self, iterations, mlambda, avg_it):
self.mlambda = float(mlambda)
self.iter = iterations
self.method = "Sarsa_control_linear_approx"
epsilon = 0.05
alpha = 0.01
l_mse = 0
e_mse = np.zeros((avg_it,self.iter))
monte_carlo_Q = pickle.load(open("Data/Qval_func_1000000_MC_control.pkl", "rb"))
n_elements = monte_carlo_Q.shape[0]*monte_carlo_Q.shape[1]*2
for my_it in xrange(avg_it):
self.Q = np.zeros((self.env.dl_values, self.env.pl_values, self.env.act_values))
self.LinE = np.zeros(len(self.d_edges)*len(self.p_edges)*2)
self.theta = np.random.random(36)*0.2
#self.theta = np.zeros(len(self.d_edges)*len(self.p_edges)*2)
count_wins = 0
# Loop over episodes (complete game runs)
for episode in xrange(self.iter):
self.LinE = np.zeros(36)
s = self.env.get_initial_state()
if np.random.random() < 1-epsilon:
Qa = -100000
a = None
for act in Actions.get_values():
phi_curr = self.feature_computation(s,act)
Q = sum(self.theta*phi_curr)
if Q > Qa:
Qa = Q
a = act
phi = phi_curr
else:
a = Actions.stick if np.random.random()<0.5 else Actions.hit
phi = self.feature_computation(s,a)
Qa = sum(self.theta*phi)
# Execute until game ends
while not s.term:
# Accumulating traces
self.LinE[phi==1] += 1
# execute action
s_next = self.env.step(s, a)
# compute delta
delta = s_next.rew - sum(self.theta*phi)
# choose next action with epsilon greedy policy
if np.random.random() < 1-epsilon:
Qa = float(-100000)
a = None
for act in Actions.get_values():
phi_curr = self.feature_computation(s_next,act)
Q = sum(self.theta*phi_curr)
if Q > Qa:
Qa = Q
a = act
phi = phi_curr
else:
a = Actions.stick if np.random.random()<0.5 else Actions.hit
phi = self.feature_computation(s_next,a)
Qa = sum(self.theta*phi)
# delta
delta += Qa
self.theta += alpha*delta*self.LinE
self.LinE = self.mlambda*self.LinE
# reassign s and a
s = s_next
#if episode%10000==0: print "Episode: %d, Reward: %d" %(episode, s_next.rew)
count_wins = count_wins+1 if s_next.rew==1 else count_wins
self.Q = self.deriveQ()
e_mse[my_it, episode] = np.sum(np.square(self.Q-monte_carlo_Q))/float(n_elements)
print float(count_wins)/self.iter*100
self.Q = self.deriveQ()
l_mse += np.sum(np.square(self.Q-monte_carlo_Q))
if mlambda==0 or mlambda==1:
plt.plot(e_mse.mean(axis=0))
plt.ylabel('mse vs episodes')
plt.show()
# Derive value function
for d in xrange(self.env.dl_values):
for p in xrange(self.env.pl_values):
self.V[d,p] = max(self.Q[d, p, :])
#print self.theta
return l_mse/float(n_elements)
def TD_control_linear(self, iterations, mlambda, avg_it):
self.mlambda = float(mlambda)
self.iter = iterations
self.method = "Sarsa_control_linear_approx"
epsilon = 0.05
alpha = 0.01
l_mse = 0
e_mse = np.zeros((avg_it, self.iter))
monte_carlo_Q = pickle.load(
open("Data/Qval_func_1000000_MC_control.pkl", "rb"))
n_elements = monte_carlo_Q.shape[0] * monte_carlo_Q.shape[1] * 2
for my_it in xrange(avg_it):
self.Q = np.zeros(
(self.env.dl_values, self.env.pl_values, self.env.act_values))
self.LinE = np.zeros(len(self.d_edges) * len(self.p_edges) * 2)
self.theta = np.random.random(36) * 0.2
#self.theta = np.zeros(len(self.d_edges)*len(self.p_edges)*2)
count_wins = 0
# Loop over episodes (complete game runs)
for episode in xrange(self.iter):
self.LinE = np.zeros(36)
s = self.env.get_initial_state()
if np.random.random() < 1 - epsilon:
Qa = -100000
a = None
for act in Actions.get_values():
phi_curr = self.feature_computation(s, act)
Q = sum(self.theta * phi_curr)
if Q > Qa:
Qa = Q
a = act
phi = phi_curr
else:
a = Actions.stick if np.random.random(
) < 0.5 else Actions.hit
phi = self.feature_computation(s, a)
Qa = sum(self.theta * phi)
# Execute until game ends
while not s.term:
# Accumulating traces
self.LinE[phi == 1] += 1
# execute action
s_next = self.env.step(s, a)
# compute delta
delta = s_next.rew - sum(self.theta * phi)
# choose next action with epsilon greedy policy
if np.random.random() < 1 - epsilon:
Qa = float(-100000)
a = None
for act in Actions.get_values():
phi_curr = self.feature_computation(s_next, act)
Q = sum(self.theta * phi_curr)
if Q > Qa:
Qa = Q
a = act
phi = phi_curr
else:
a = Actions.stick if np.random.random(
) < 0.5 else Actions.hit
phi = self.feature_computation(s_next, a)
Qa = sum(self.theta * phi)
# delta
delta += Qa
self.theta += alpha * delta * self.LinE
self.LinE = self.mlambda * self.LinE
# reassign s and a
s = s_next
#if episode%10000==0: print "Episode: %d, Reward: %d" %(episode, s_next.rew)
count_wins = count_wins + 1 if s_next.rew == 1 else count_wins
self.Q = self.deriveQ()
e_mse[my_it, episode] = np.sum(
np.square(self.Q - monte_carlo_Q)) / float(n_elements)
print float(count_wins) / self.iter * 100
self.Q = self.deriveQ()
l_mse += np.sum(np.square(self.Q - monte_carlo_Q))
if mlambda == 0 or mlambda == 1:
plt.plot(e_mse.mean(axis=0))
plt.ylabel('mse vs episodes')
plt.show()
# Derive value function
for d in xrange(self.env.dl_values):
for p in xrange(self.env.pl_values):
self.V[d, p] = max(self.Q[d, p, :])
#print self.theta
return l_mse / float(n_elements)