def MC_control(self, iterations):
# Initialise
self.iter = iterations
self.method = "MC_control"
count_wins = 0
episode_pairs = []
# Loop over episodes (complete game runs)
for episode in xrange(self.iter):
# reset state action pair list
episode_pairs = []
# get initial state for current episode
my_state = self.env.get_initial_state()
# Execute until game ends
while not my_state.term:
# choose action with epsilon greedy policy
my_action = self.eps_greedy_choice(my_state)
# store action state pairs
episode_pairs.append((my_state, my_action))
# update visits
self.N[my_state.dl_sum - 1, my_state.pl_sum - 1,
Actions.get_value(my_action)] += 1
# execute action
my_state = self.env.step(my_state, my_action)
# Update Action value function accordingly
for curr_s, curr_a in episode_pairs:
step = 1.0 / (self.N[curr_s.dl_sum - 1, curr_s.pl_sum - 1,
Actions.get_value(curr_a)])
error = my_state.rew - self.Q[curr_s.dl_sum - 1,
curr_s.pl_sum - 1,
Actions.get_value(curr_a)]
self.Q[curr_s.dl_sum - 1, curr_s.pl_sum - 1,
Actions.get_value(curr_a)] += step * error
#if episode%10000==0: print "Episode: %d, Reward: %d" %(episode, my_state.rew)
count_wins = count_wins + 1 if my_state.rew == 1 else count_wins
print float(count_wins) / self.iter * 100
# 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, :])
def MC_control(self, iterations):
# Initialise
self.iter = iterations
self.method = "MC_control"
count_wins = 0
episode_pairs = []
# Loop over episodes (complete game runs)
for episode in xrange(self.iter):
# reset state action pair list
episode_pairs = []
# get initial state for current episode
my_state = self.env.get_initial_state()
# Execute until game ends
while not my_state.term:
# choose action with epsilon greedy policy
my_action = self.eps_greedy_choice(my_state)
# store action state pairs
episode_pairs.append((my_state, my_action))
# update visits
self.N[my_state.dl_sum-1, my_state.pl_sum-1, Actions.get_value(my_action)] += 1
# execute action
my_state = self.env.step(my_state, my_action)
# Update Action value function accordingly
for curr_s, curr_a in episode_pairs:
step = 1.0 / (self.N[curr_s.dl_sum-1, curr_s.pl_sum-1, Actions.get_value(curr_a)])
error = my_state.rew - self.Q[curr_s.dl_sum-1, curr_s.pl_sum-1, Actions.get_value(curr_a)]
self.Q[curr_s.dl_sum-1, curr_s.pl_sum-1, Actions.get_value(curr_a)] += step * error
#if episode%10000==0: print "Episode: %d, Reward: %d" %(episode, my_state.rew)
count_wins = count_wins+1 if my_state.rew==1 else count_wins
print float(count_wins)/self.iter*100
# 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, :])
def TD_control(self, iterations, mlambda, avg_it):
self.mlambda = float(mlambda)
self.iter = iterations
self.method = "Sarsa_control"
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.N = np.zeros((self.env.dl_values, self.env.pl_values, self.env.act_values))
self.Q = np.zeros((self.env.dl_values, self.env.pl_values, self.env.act_values))
self.E = np.zeros((self.env.dl_values, self.env.pl_values, self.env.act_values))
count_wins = 0
# Loop over episodes (complete game runs)
for episode in xrange(self.iter):
self.E = np.zeros((self.env.dl_values, self.env.pl_values, self.env.act_values))
s = self.env.get_initial_state()
a = self.eps_greedy_choice(s)
# Execute until game ends
while not s.term:
# update visit count
self.N[s.dl_sum-1, s.pl_sum-1, Actions.get_value(a)] += 1
# execute action
s_next = self.env.step(s, a)
# choose next action with epsilon greedy policy
a_next = self.eps_greedy_choice(s_next)
# update action value function
alpha = 1.0 / (self.N[s.dl_sum-1, s.pl_sum-1, Actions.get_value(a)])
try:
delta = s_next.rew + self.Q[s_next.dl_sum-1, s_next.pl_sum-1, Actions.get_value(a_next)] -\
self.Q[s.dl_sum-1, s.pl_sum-1, Actions.get_value(a)]
except:
delta = s_next.rew - self.Q[s.dl_sum-1, s.pl_sum-1, Actions.get_value(a)]
self.E[s.dl_sum-1, s.pl_sum-1, Actions.get_value(a)] += 1
update = alpha*delta*self.E
self.Q = self.Q+update
self.E = self.mlambda*self.E
# reassign s and a
s = s_next
a = a_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
e_mse[my_it, episode] = np.sum(np.square(self.Q-monte_carlo_Q))/float(n_elements)
print float(count_wins)/self.iter*100
l_mse += np.sum(np.square(self.Q-monte_carlo_Q))/float(n_elements)
#print n_elements
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, :])
return float(l_mse)/avg_it
def TD_control(self, iterations, mlambda, avg_it):
self.mlambda = float(mlambda)
self.iter = iterations
self.method = "Sarsa_control"
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.N = np.zeros(
(self.env.dl_values, self.env.pl_values, self.env.act_values))
self.Q = np.zeros(
(self.env.dl_values, self.env.pl_values, self.env.act_values))
self.E = np.zeros(
(self.env.dl_values, self.env.pl_values, self.env.act_values))
count_wins = 0
# Loop over episodes (complete game runs)
for episode in xrange(self.iter):
self.E = np.zeros((self.env.dl_values, self.env.pl_values,
self.env.act_values))
s = self.env.get_initial_state()
a = self.eps_greedy_choice(s)
# Execute until game ends
while not s.term:
# update visit count
self.N[s.dl_sum - 1, s.pl_sum - 1,
Actions.get_value(a)] += 1
# execute action
s_next = self.env.step(s, a)
# choose next action with epsilon greedy policy
a_next = self.eps_greedy_choice(s_next)
# update action value function
alpha = 1.0 / (self.N[s.dl_sum - 1, s.pl_sum - 1,
Actions.get_value(a)])
try:
delta = s_next.rew + self.Q[s_next.dl_sum-1, s_next.pl_sum-1, Actions.get_value(a_next)] -\
self.Q[s.dl_sum-1, s.pl_sum-1, Actions.get_value(a)]
except:
delta = s_next.rew - self.Q[s.dl_sum - 1, s.pl_sum - 1,
Actions.get_value(a)]
self.E[s.dl_sum - 1, s.pl_sum - 1,
Actions.get_value(a)] += 1
update = alpha * delta * self.E
self.Q = self.Q + update
self.E = self.mlambda * self.E
# reassign s and a
s = s_next
a = a_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
e_mse[my_it, episode] = np.sum(
np.square(self.Q - monte_carlo_Q)) / float(n_elements)
print float(count_wins) / self.iter * 100
l_mse += np.sum(
np.square(self.Q - monte_carlo_Q)) / float(n_elements)
#print n_elements
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, :])
return float(l_mse) / avg_it