def execute_episode():
history = []
for state in range(N * M):
action = e_greedy(state,
get_valid_actions(maze, state, ACTIONS),
Q,
epsilon=0.1)
history.append((state, action))
R = reward(maze, state)
next_state = get_next_state(maze, state, action)
next_best = max(q_table[next_state])
td_target = R + DISCOUNT * next_best
td_diff = td_target - q_table[state][action]
q_table[state][action] += LEARNING_RATE * td_diff
q_table[state][action] = round(q_table[state][action], 3)
model[state][action] = (next_state, R)
for _ in range(PLANNING):
state, action = random.choice(history)
next_state, R = model[state][action]
next_best = max(q_table[next_state])
td_target = R + DISCOUNT * next_best
td_diff = td_target - q_table[state][action]
q_table[state][action] += LEARNING_RATE * td_diff
q_table[state][action] = round(q_table[state][action], 3)
def get_path():
i = 0
path = [i]
while i != N * M - 1:
i = get_next_state(maze, i, ACTIONS[np.argmax(q_table[i])])
path.append(i)
return path
def reward(maze, state, action):
next_state = get_next_state(maze, state, action)
x, y = state_to_cell(maze, state)
if maze.grid[x][y].neighbors[ACTION_MAPPER[action]] == INF:
return -10000
elif next_state == (maze.num_rows * maze.num_columns - 1):
return 1000
else:
return 0
def execute_episode():
for state in range(N * M):
action = e_greedy(state,
get_valid_actions(maze, state, ACTIONS),
Q,
epsilon=0.1)
R = reward(maze, state)
next_state = get_next_state(maze, state, action)
next_best = max(q_table[next_state])
td_target = R + DISCOUNT * next_best
td_diff = td_target - q_table[state][action]
q_table[state][action] += LEARNING_RATE * td_diff
q_table[state][action] = round(q_table[state][action], 3)
def get_path():
i = N * M - 2
path = [i]
visited = set()
while i:
temp = []
for action, value in enumerate(q_table[i]):
temp.append((value, action))
temp.sort()
for item in temp:
try:
new_state = get_next_state(maze, i, item[1])
if new_state not in visited:
i = new_state
path.append(i)
break
except:
continue
return path