def learn(self, episodes, visualize=False):
'''
Learn best Q-values
:param episodes: Number of episodes to learn
:param visualize: if True, render environment when learning
'''
for ep in range(episodes):
state = self.env.reset()
t = 0
while True:
if visualize:
self.env.render()
action = self.act(discretize(state), episode=ep)
'''
TODO:
1) Run the enviroment with chosen action (action)
2) call update Qs (don't forget to discretize states)
3) S = S'
'''
raise NotImplementedError('Please implement the loop body')
t += 1
if done:
print("Episode {} finished after {} timesteps".format(
ep, t))
break
def learn(self, episodes, visualize=False):
'''
Learn best Q-values
:param episodes: Number of episodes to learn
:param visualize: if True, render environment when learning
'''
for ep in range(episodes):
state = self.env.reset()
t = 0
while True:
if visualize:
self.env.render()
action = self.act(discretize(state), episode=ep)
new_state, reward, done, info = self.env.step(action)
self.updateQ(discretize(state), action, reward,
discretize(new_state))
state = new_state
t += 1
if done:
print("Episode {} finished after {} timesteps".format(
ep, t))
break
def edgeLink(M, Mag, Ori):
# Multiply magnitude with
binary_mag = np.multiply(M, Mag)
row, column = binary_mag.shape
edge_map = np.empty([row, column])
# Declare low and high threshold
threshold_low = 0.03 * np.amax(binary_mag)
threshold_high = 2.5 * threshold_low
# Set thresholds
for j in range(0, column):
for i in range(0, row):
if binary_mag[i, j] > threshold_high:
edge_map[i, j] = 1
elif binary_mag[i, j] <= threshold_low:
edge_map[i, j] = 0
else:
edge_map[i, j] = 0.5
# Hysteresis
for j in range(0, column):
for i in range(0, row):
# Start at strong edge
if edge_map[i, j] == 0.5:
pi = math.pi
# Orientation of gradient
angle = Ori[i, j]
# Orientations of edge perpendicular to gradient direction
angle_neg = discretize(angle - np.cos(angle))
angle_pos = discretize(angle + np.cos(angle))
neighbor_1 = 0.0
neighbor_2 = 0.0
row_1 = 0
row_2 = 0
column_1 = 0
column_2 = 0
if angle_neg == 0 or angle_neg == pi or angle_pos == 0 or angle_pos == pi:
if j + 1 < column:
neighbor_1 = edge_map[i][j + 1]
row_1 = i
column_1 = j + 1
if j - 1 >= 0:
neighbor_2 = edge_map[i][j - 1]
row_2 = i
column_2 = j - 1
elif angle_neg == pi / 2 or angle_neg == 3 * pi / 2 or angle_pos == pi / 2 or angle_pos == 3 * pi / 2:
if i + 1 < row:
neighbor_1 = edge_map[i + 1][j]
row_1 = i + 1
column_1 = j
if i - 1 >= 0:
neighbor_2 = edge_map[i - 1][j]
row_2 = i - 1
column_2 = j
elif angle_neg == pi / 4 or angle_neg == 5 * pi / 4 or angle_pos == pi / 4 or angle_pos == 5 * pi / 4:
if i + 1 < row and j + 1 < column:
neighbor_1 = edge_map[i + 1][j + 1]
row_1 = i + 1
column_1 = j + 1
if i - 1 >= 0 and j - 1 >= 0:
neighbor_2 = edge_map[i - 1][j - 1]
row_2 = i - 1
column_2 = j - 1
elif angle_neg == 3 * pi / 4 or angle_neg == 7 * pi / 4 or angle_pos == 3 * pi / 4 or angle_pos == 7 * pi / 4:
if i - 1 >= 0 and j + 1 < column:
neighbor_1 = edge_map[i - 1][j + 1]
row_1 = i - 1
column_1 = j + 1
if i + 1 < row and j - 1 >= 0:
neighbor_2 = edge_map[i + 1][j - 1]
row_2 = i + 1
column_2 = j - 1
ori_difference_1 = abs(Ori[i][j] - Ori[row_1][column_1])
ori_difference_2 = abs(Ori[i][j] - Ori[row_2][column_2])
if neighbor_1 < 1 and neighbor_2 < 1:
edge_map[i][j] = 0
elif neighbor_1 == 1 and neighbor_2 < 1:
if ori_difference_1 >= threshold_low and ori_difference_1 <= threshold_high:
edge_map[i][j] = 1
else:
edge_map[i][j] = 0
elif neighbor_2 == 1 and neighbor_1 < 1:
if ori_difference_2 >= threshold_low and ori_difference_2 <= threshold_high:
edge_map[i][j] = 1
else:
edge_map[i][j] = 0
elif neighbor_1 == 1 and neighbor_2 == 1:
if ori_difference_1 >= threshold_low and ori_difference_1 <= threshold_high or ori_difference_2 >= threshold_low and ori_difference_2 <= threshold_high:
edge_map[i][j] = 1
else:
edge_map[i][j] = 0
# Final for loop to clear weak values
for j in range(0, column):
for i in range(0, row):
if edge_map[i][j] == 0.5:
edge_map[i][j] = 0
return edge_map
def preprocess_state(self, state):
"""Map a continuous state to its discretized representation."""
return tuple(discretize(state,self.state_grid))
TRIALS = 10
######## CONSTANTS ########
env = gym.make('MountainCar-v0')
agent = QLAgent(env=env,
n_states=N_STATES,
n_actions=N_ACTIONS,
lr=0.5,
discount=0.95)
# Train agent
agent.learn(episodes=EPISODES)
# Evaluate
success = 0
for tr in range(TRIALS):
state = env.reset()
t = 0
while True:
env.render()
action = agent.act(discretize(state))
state, reward, done, info = env.step(action)
t += 1
if done:
print("Trial {} finished after {} timesteps".format(tr, t))
if t < 200:
success += 1
break
print("Success: %d/%d" % (success, TRIALS))
env.close()