def main():
max_episodes = 1500
for episode in range(max_episodes):
done = False
step_count = 0
env.reset()
while not done:
state, reward, done = env.step(env.get_random_actions()[0])
step_count += 1
def main():
max_episodes = 1500
for episode in range(max_episodes):
done = False
step_count = 0
env.reset()
obs, _, _, _, _, _, _, _, _ = env.step(0)
cv2.imshow('mario', obs)
cv2.waitKey(0)
cv2.destroyAllWindows()
while not done:
state, reward, done, s4, s5, s6, s7, r_d, s_d = env.step(11)
# 0 next_state:{ndarray} shape (90,90)
# 1 reward:{int}
# 2 done:{bool}
# 3 state_clear:{bool}
# 4 max_x:{int}
# 5 time_out:{bool}
# 6 now_x:{int}
step_count += 1
def learn(env, MAX_EPISODE, EPS_START, EPS_END, EPS_DECAY, ACTION_NUM,
REPLAY_MEMORY_CAPACITY, BATCH_SIZE, LOSS_FUNCTION, OPTIM_METHOD,
LEARNING_RATE, GAMMA, NET_COPY_STEP, OBSERVE, TRAIN_FREQ, PATH):
### initialization ###
action_space = [(0, Tensor([1, 0, 0, 0, 0, 0])),
(3, Tensor([0, 1, 0, 0, 0, 0])),
(7, Tensor([0, 0, 1, 0, 0, 0])),
(11, Tensor([0, 0, 0, 1, 0, 0])),
(4, Tensor([0, 0, 0, 0, 1, 0])),
(8, Tensor([0, 0, 0, 0, 0, 1]))]
# (action_button , action_onehot)
# 以上动作分别为：不动、左走、右走、跳、左跳、右跳
value_net = dqn_net(ACTION_NUM)
target_net = dqn_net(ACTION_NUM)
if torch.cuda.is_available():
value_net.cuda()
target_net.cuda()
if os.path.isfile(PATH):
value_net.load_state_dict(torch.load(PATH))
buffer = replay_memory(REPLAY_MEMORY_CAPACITY)
env.reset()
obs, _, _, _, _, _, _ = env.step(0)
obs = ob_process(obs)
obs4 = torch.cat(([obs, obs, obs, obs]),
dim=0) # {Tensor} of shape torch.Size([4,84,84])
judge_distance = 0
episode_total_reward = 0
epi_total_reward_list = []
mean_reward_list = []
# counters #
time_step = 0
update_times = 0
episode_num = 0
history_distance = 200
while episode_num <= MAX_EPISODE:
### choose an action with epsilon-greedy ###
prob = random.random()
threshold = EPS_END + (EPS_START - EPS_END) * math.exp(
-1 * episode_num / EPS_DECAY)
if prob <= threshold:
action_index = np.random.randint(6)
action_button = action_space[action_index][0] # {int}
action_onehot = action_space[action_index][1] # {Tensor}
else:
action_button, action_onehot = value_net.select_action(obs4)
### do one step ###
obs_next, reward, done, _, max_distance, _, now_distance = env.step(
action_button)
obs_next = ob_process(obs_next)
obs4_next = torch.cat(([obs4[1:, :, :], obs_next]), dim=0)
buffer.add(obs4.unsqueeze(0), action_onehot.unsqueeze(0),
obs4_next.unsqueeze(0),
Tensor([reward]).unsqueeze(0), done)
episode_total_reward += reward
if now_distance <= history_distance:
judge_distance += 1
else:
judge_distance = 0
history_distance = max_distance
'''the transition added to buffer
obs4: {ndarray} size (4,84,84)
action: {list} size 6 e.g. [1,0,0,0,0,0] one hot list
obs_next: {ndarray} size (84,84)
reward: {int}
done: {bool}
'''
### go to the next state ###
if done == False:
obs4 = obs4_next
time_step += 1
elif done == True or judge_distance > 50:
env.reset()
obs, _, _, _, _, _, _ = env.step(0)
obs = ob_process(obs)
obs4 = torch.cat(([obs, obs, obs, obs]), dim=0)
episode_num += 1
history_distance = 200
# plot graph #
epi_total_reward_list.append(episode_total_reward)
mean100 = np.mean(epi_total_reward_list[-101:-1])
mean_reward_list.append(mean100)
plot_graph(mean_reward_list)
print('episode %d total reward=%.2f' %
(episode_num, episode_total_reward))
episode_total_reward = 0
### do one step update ###
if time_step >= OBSERVE and time_step % TRAIN_FREQ == 0:
batch_transition = buffer.sample(BATCH_SIZE)
'''{Transition}
0:{tuple} of {Tensor}-shape-torch.Size([4,84,84])
1:{tuple} of {Tensor}-shape-torch.Size([6])
2:{tuple} of {Tensor}-shape-torch.Size([4,84,84])
3:{tuple} of {int}
4:{tuple} of {bool}
'''
value_net.update(samples=batch_transition,
loss_func=LOSS_FUNCTION,
optim_func=OPTIM_METHOD,
learn_rate=LEARNING_RATE,
target_net=target_net,
BATCH_SIZE=BATCH_SIZE,
GAMMA=GAMMA)
update_times += 1
### copy value net parameters to target net ###
if update_times % NET_COPY_STEP == 0:
target_net.load_state_dict(value_net.state_dict())
torch.save(value_net.state_dict(), PATH)
def learn(env,
MAX_EPISODE,
EPS_START,
EPS_END,
EPS_DECAY,
LEARNING_RATE,
GAMMA,
):
mapping_reduced_action = [3, 7, 11, 4, 10]
Un_mapping_reduced_action = [100, 100, 100, 0, 3, 100, 100, 1, 100, 100, 4, 2]
### initialization ###
env.reset()
obs, _, _, _, _, _, _, _, _, _ = env.step(0)
judge_distance = 0
episode_total_reward = 0
no_states_observed = 1
epi_total_reward_list = []
mean_reward_list = []
filename = 'State_Q_Table.csv'
LEARNING_RATE_CTR = [np.zeros(6), np.zeros(6)]
try:
state_table = np.loadtxt(filename, delimiter=",", usecols=[0])
state_table = state_table.astype(int)
state_table = state_table.tolist()
q_table = np.loadtxt(filename, delimiter=",", usecols=[1, 2, 3, 4, 5, 6])
LEARNING_RATE_CTR = q_table * 0
q_table = q_table.tolist()
LEARNING_RATE_CTR = LEARNING_RATE_CTR.tolist()
no_states_observed = len(state_table) - 1
except:
print('warning: Error %s: Loading State, Action Table' % filename)
state_table = [0, 1]
q_table = [np.random.rand(6), np.random.rand(6)]
if (state_table == [] or q_table == []):
state_table = [0, 1]
q_table = [np.random.rand(6), np.random.rand(6)]
LEARNING_RATE_CTR = [np.zeros(6), np.zeros(6)]
# counters #
time_step = 0
update_times = 0
episode_num = 0
history_distance = 200
index_s = 0
state_d_current = state_table[index_s]
f_handle = open(filename, 'w')
f_handle_Evo = open('State_Q_Table_ev.csv', 'a')
controller_speed_ctr = 0
reward_collection = 0
print(state_table)
print(q_table)
while episode_num <= MAX_EPISODE:
### choose an action with epsilon-greedy ###
prob = random.random()
threshold = EPS_END # + (EPS_START - EPS_END) * math.exp(-1 * episode_num / EPS_DECAY)
# action_onehot = action_space[0][1] # {Tensor}
# if(controller_speed_ctr == 0):
reward_collection = 0
if prob <= threshold:
action_button_d = np.random.randint(6)
else:
action_button_d = np.argmax(q_table[index_s])
np.savetxt(f_handle_Evo, [np.concatenate([[state_d_current, action_button_d], q_table[index_s]])], fmt='%1.6f',
delimiter=',')
obs_next, reward, done, _, max_distance, _, now_distance, reward_d, state_d_next, keyboard_keys = env.step(
action_button_d)
reward_collection += reward_d
obs_next, reward, done, _, max_distance, _, now_distance, reward_d, state_d_next, keyboard_keys = env.step(
action_button_d)
reward_collection += reward_d
obs_next, reward, done, _, max_distance, _, now_distance, reward_d, state_d_next, keyboard_keys = env.step(
action_button_d)
reward_collection += reward_d
obs_next, reward, done, _, max_distance, _, now_distance, reward_d, state_d_next, keyboard_keys = env.step(
action_button_d)
reward_collection += reward_d
if state_d_next in state_table:
# start = time.clock()
next_index = state_table.index(state_d_next)
current_index = state_table.index(state_d_current)
current_value = q_table[current_index][action_button_d]
LEARNING_RATE_CTR[current_index][action_button_d] += 1
LEARNING_RATE_S_A = LEARNING_RATE / LEARNING_RATE_CTR[current_index][action_button_d]
# print(LEARNING_RATE_S_A)
q_table[current_index][action_button_d] = current_value + LEARNING_RATE_S_A * (
reward_d + GAMMA * (max(q_table[next_index])) - current_value)
# print(current_value + LEARNING_RATE * (reward_d + GAMMA*(max(q_table[next_index])) - current_value))
index_s = next_index
# print(q_table[current_index])
# print(current_index)
# print(time.clock() - start)
# print(np.concatenate(([state_table[current_index]], [reward_d], q_table[current_index])))
else:
state_table.append(state_d_next)
q_table.append(np.random.rand(6))
LEARNING_RATE_CTR.append(np.zeros(6))
no_states_observed = len(state_table) - 1 # no_states_observed + 1
index_s = no_states_observed
# print(no_states_observed)
print(np.concatenate(([state_table[no_states_observed]], [reward_d], q_table[no_states_observed])))
state_d_current = state_d_next
episode_total_reward += reward_d
if now_distance <= history_distance:
judge_distance += 1
else:
judge_distance = 0
history_distance = max_distance
### go to the next state ###
if done == False:
# obs4 = obs4_next
time_step += 1
elif done == True or judge_distance > 50:
env.reset()
obs, _, _, _, _, _, _, _, _, _ = env.step(0)
episode_num += 1
history_distance = 200
# plot graph #
epi_total_reward_list.append(episode_total_reward)
mean100 = np.mean(epi_total_reward_list[-101:-1])
mean_reward_list.append(mean100)
plot_graph(epi_total_reward_list)
print('episode %d total reward=%.2f' % (episode_num, episode_total_reward))
episode_total_reward = 0
np.savetxt(f_handle, np.column_stack((state_table, q_table)), fmt=','.join(['%i'] + ['%1.6f'] * 6), delimiter=',')
np.savetxt('Reward.csv', epi_total_reward_list, fmt='%1.6f')
np.savetxt('LR_CTR.csv', LEARNING_RATE_CTR, fmt='%i', delimiter=',')
f_handle.close()
f_handle_Evo.close()