def step(self, action):
self.state = self.state + np.asarray(action)
env = CartPoleEnv(self.state[0], self.state[1], self.state[2],
self.state[3], self.state[4])
episode_count = len(self.action_record)
model_diff = 0
for i in range(episode_count):
ob = env.reset()
traj_state = []
for j in range(len(self.action_record[i])):
# The traj that done is better tricky here
action = self.action_record[i][j]
ob, reward, done, _ = env.step(action)
traj_state.append(ob)
if done:
break
if not done:
model_diff = model_diff + 1 # penalty for not done
model_diff = model_diff + self._traj_diff(np.asarray(traj_state),
self.state_record[i])
reward = -model_diff - self.status
self.status = -model_diff
done = False
return np.array(self.state), reward, done, {}
def looping(qt=None, epsilon=config.epsilon, visu=False):
plt.ion()
cart = CartPoleEnv()
data = []
data_rm = []
if (qt is None):
qt = initialize_Qtable()
for episode in range(config.episodes):
cart.reset()
turn = 0
end = False
epsilon = epsilon * 0.9999
while not end:
current_state = cart.state
action = choose_action(current_state, qt, epsilon)
new_state, reward, end, _ = cart.step(action)
if end:
reward = -10
update_qt_new(qt, current_state, reward, action, new_state)
turn += 1
if (visu):
cart.render()
data.append(turn)
data_rm.append(np.mean(data[-100:]))
print("Episode: ", episode, "\tTurn:", turn, "\t Epsilon:", epsilon)
if episode % config.graph_update == 0 and episode != 0:
graph(data, data_rm)
# if ((episode + 1) % 100 == 0 and input("continue (y/n)" != "y")):
# break
cart.close()
return (data, qt)
def loop(qt=None, epsilon=1, visu=False):
plt.ion()
cart = CartPoleEnv()
data = []
data_rm = []
config.epsilon = epsilon
if (qt is None):
qt = initialize_Qtable()
for episode in range(config.episodes):
cart.reset()
turn = 0
s = cart.state
end = False
epsilon_tmp = config.epsilon
while not end:
config.epsilon *= 0.97
if (visu):
cart.render()
a = choose_action(s, qt)
_, _, end, _ = cart.step(a)
l_val = bellman_q(s, qt, dummy_cart(s), action=a)
# print(l_val)
update_qt(qt, s, a, l_val)
s = cart.state
turn += 1
data.append(turn)
data_rm.append(np.mean(data[-100:]))
print("Episode: ", episode, "\tTurn:", turn, "\t Epsilon:",
config.epsilon)
config.epsilon = epsilon_tmp
if episode % config.graph_update == 0 and episode != 0:
graph(data, data_rm)
# if ((episode + 1) % 100 == 0 and input("continue (y/n)" != "y")):
# break
cart.close()
return (data, qt)
V = np.zeros(n_s)
for i in range(episodes):
print("episode=",i)
obs = env.reset()
s = discretize_state(obs,s_bounds,n_s)
finished = False
time_step=0
#dlugie epizody przerywamy po 200 krokach
while not finished and not time_step==200:
#polityka stochastyczna - prawdopodobienstwo 0.5 dla obu akcji (w lewo, w prawo)
action = np.random.randint(0,2)
obs, reward, finished, info = env.step(action)
state_new = discretize_state(obs,s_bounds,n_s)
V[s] = V[s] + alpha * (reward + gamma*V[state_new] - V[s])
s = state_new
time_step+=1
#DO UZUPEŁNIENIA
print(V)
state = env.reset()
state = np.reshape(state, [1, 4]) # reshape from [[a, b]] to [a, b]
for t in range(1000):
action, force = agent.act(state)
# given Q values for each force and max arg of it, apply the corresponding force to the model
if force == 2:
env.force_mag = 6
elif force == 3:
env.force_mag = 8
else:
env.force_mag = 10
# perform one step
state_, reward, done, info = env.step(action)
state_ = np.reshape(state_, [1, 4])
# reward function
reward = reward if not done else -20 # additional penalty for a loss
reward -= 1.0 * abs(
state_[0]
[0]) # penalty for moving too far away, increasing linearly
# archive the step and perform fitting to the model
agent.remember(state, reward, action, state_, done)
# agent.replay()
# state = next state
state = state_
from cartpole import CartPoleEnv
import numpy as np
cart = CartPoleEnv()
cart.reset()
for _ in range(1000):
# Calculate the Gradients
# Update Thetas
# Sample u trajectory
# Apply u[0] to the actual system
cart.step(10) # Apply Some force
# Update the New State in the Learner
# Shift the Thetas
# Simulate
cart.render()
cart.close()
def main():
# Define dimensions of the networks
meta_value_input_dim = STATE_DIM + TASK_CONFIG_DIM # 7
task_config_input_dim = STATE_DIM + ACTION_DIM + 1 # 7
# init meta value network with a task config network
meta_value_network = MetaValueNetwork(input_size = meta_value_input_dim,hidden_size = 80,output_size = 1)
task_config_network = TaskConfigNetwork(input_size = task_config_input_dim,hidden_size = 30,num_layers = 1,output_size = 3)
meta_value_network.cuda()
task_config_network.cuda()
if os.path.exists("meta_value_network_cartpole.pkl"):
meta_value_network.load_state_dict(torch.load("meta_value_network_cartpole.pkl"))
print("load meta value network success")
if os.path.exists("task_config_network_cartpole.pkl"):
task_config_network.load_state_dict(torch.load("task_config_network_cartpole.pkl"))
print("load task config network success")
meta_value_network_optim = torch.optim.Adam(meta_value_network.parameters(),lr=0.001)
task_config_network_optim = torch.optim.Adam(task_config_network.parameters(),lr=0.001)
# init a task generator for data fetching
task_list = [CartPoleEnv(np.random.uniform(L_MIN,L_MAX)) for task in range(TASK_NUMS)]
[task.reset() for task in task_list]
task_lengths = [task.length for task in task_list]
print("task length:",task_lengths)
for episode in range(EPISODE):
# ----------------- Training ------------------
if (episode+1) % 10 ==0 :
# renew the tasks
task_list = [CartPoleEnv(np.random.uniform(L_MIN,L_MAX)) for task in range(TASK_NUMS)]
task_lengths = [task.length for task in task_list]
print("task length:",task_lengths)
[task.reset() for task in task_list]
# fetch pre data samples for task config network
# [task_nums,sample_nums,x+y`]
actor_network_list = [ActorNetwork(STATE_DIM,40,ACTION_DIM) for i in range(TASK_NUMS)]
[actor_network.cuda() for actor_network in actor_network_list]
actor_network_optim_list = [torch.optim.Adam(actor_network.parameters(),lr = 0.01) for actor_network in actor_network_list]
# sample pre state,action,reward for task config
pre_states = []
pre_actions = []
pre_rewards = []
for i in range(TASK_NUMS):
states,actions,rewards,_,_ = roll_out(actor_network_list[i],task_list[i],SAMPLE_NUMS)
pre_states.append(states)
pre_actions.append(actions)
pre_rewards.append(rewards)
for step in range(STEP):
for i in range(TASK_NUMS):
# init task config [1, sample_nums,task_config] task_config size=3
pre_data_samples = torch.cat((pre_states[i][-9:],pre_actions[i][-9:],torch.Tensor(pre_rewards[i])[-9:]),1).unsqueeze(0)
task_config = task_config_network(Variable(pre_data_samples).cuda()) # [1,3]
states,actions,rewards,is_done,final_state = roll_out(actor_network_list[i],task_list[i],SAMPLE_NUMS)
final_r = 0
if not is_done:
value_inputs = torch.cat((Variable(final_state.unsqueeze(0)).cuda(),task_config.detach()),1)
final_r = meta_value_network(value_inputs).cpu().data.numpy()[0]
# train actor network
actor_network_optim_list[i].zero_grad()
states_var = Variable(states).cuda()
actions_var = Variable(actions).cuda()
task_configs = task_config.repeat(1,len(rewards)).view(-1,3)
log_softmax_actions = actor_network_list[i](states_var)
vs = meta_value_network(torch.cat((states_var,task_configs.detach()),1)).detach()
# calculate qs
qs = Variable(torch.Tensor(discount_reward(rewards,0.99,final_r))).cuda()
advantages = qs - vs
actor_network_loss = - torch.mean(torch.sum(log_softmax_actions*actions_var,1)* advantages) #+ entropy #+ actor_criterion(actor_y_samples,target_y)
actor_network_loss.backward()
torch.nn.utils.clip_grad_norm(actor_network_list[i].parameters(),0.5)
actor_network_optim_list[i].step()
# train value network
meta_value_network_optim.zero_grad()
target_values = qs
values = meta_value_network(torch.cat((states_var,task_configs),1))
criterion = nn.MSELoss()
meta_value_network_loss = criterion(values,target_values)
meta_value_network_loss.backward()
torch.nn.utils.clip_grad_norm(meta_value_network.parameters(),0.5)
meta_value_network_optim.step()
# train actor network
pre_states[i] = states
pre_actions[i] = actions
pre_rewards[i] = rewards
if (step + 1) % 100 == 0:
result = 0
test_task = CartPoleEnv(length = task_list[i].length)
for test_epi in range(10):
state = test_task.reset()
for test_step in range(200):
softmax_action = torch.exp(actor_network_list[i](Variable(torch.Tensor([state])).cuda()))
#print(softmax_action.data)
action = np.argmax(softmax_action.cpu().data.numpy()[0])
next_state,reward,done,_ = test_task.step(action)
result += reward
state = next_state
if done:
break
print("episode:",episode,"task:",i,"step:",step+1,"test result:",result/10.0)
if (episode+1) % 10 == 0 :
# Save meta value network
torch.save(meta_value_network.state_dict(),"meta_value_network_cartpole.pkl")
torch.save(task_config_network.state_dict(),"task_config_network_cartpole.pkl")
print("save networks for episode:",episode)
from cartpole import CartPoleEnv
env = CartPoleEnv(length=1.0)
env.reset()
for step in range(1000):
action = 0
next_state, reward, done, _ = env.step(0)
if done:
print "done reward:", reward
break
def main():
# Define dimensions of the networks
meta_value_input_dim = STATE_DIM + TASK_CONFIG_DIM # 7
task_config_input_dim = STATE_DIM + ACTION_DIM + 1 # 7
# init meta value network with a task config network
meta_value_network = MetaValueNetwork(input_size = meta_value_input_dim,hidden_size = 80,output_size = 1)
task_config_network = TaskConfigNetwork(input_size = task_config_input_dim,hidden_size = 30,num_layers = 1,output_size = 3)
meta_value_network.cuda()
task_config_network.cuda()
if os.path.exists("meta_value_network_cartpole.pkl"):
meta_value_network.load_state_dict(torch.load("meta_value_network_cartpole.pkl"))
print("load meta value network success")
if os.path.exists("task_config_network_cartpole.pkl"):
task_config_network.load_state_dict(torch.load("task_config_network_cartpole.pkl"))
print("load task config network success")
task_lengths = np.linspace(L_MIN,L_MAX,TASK_NUMS)
datas = []
for task_length in task_lengths:
data_i = {}
data_i["task_length"] = task_length
data_i_episode = {}
for episode in range(EPISODE):
task = CartPoleEnv(length = task_length)
task.reset()
data_i_episode["episode"] = episode
# ----------------- Training ------------------
# fetch pre data samples for task config network
# [task_nums,sample_nums,x+y`]
actor_network = ActorNetwork(STATE_DIM,40,ACTION_DIM)
actor_network.cuda()
actor_network_optim = torch.optim.Adam(actor_network.parameters(),lr = 0.01)
'''
if os.path.exists("actor_network.pkl"):
actor_network.load_state_dict(torch.load("actor_network.pkl"))
print("load actor_network success")
'''
# sample pre state,action,reward for task confi
pre_states,pre_actions,pre_rewards,_,_ = roll_out(actor_network,task,SAMPLE_NUMS)
test_results = []
train_games = []
for step in range(STEP):
# init task config [1, sample_nums,task_config] task_config size=3
pre_data_samples = torch.cat((pre_states[-9:],pre_actions[-9:],torch.Tensor(pre_rewards)[-9:]),1).unsqueeze(0)
task_config = task_config_network(Variable(pre_data_samples).cuda()) # [1,3]
states,actions,rewards,is_done,final_state = roll_out(actor_network,task,SAMPLE_NUMS)
final_r = 0
if not is_done:
value_inputs = torch.cat((Variable(final_state.unsqueeze(0)).cuda(),task_config.detach()),1)
final_r = meta_value_network(value_inputs).cpu().data.numpy()[0]
# train actor network
actor_network_optim.zero_grad()
states_var = Variable(states).cuda()
actions_var = Variable(actions).cuda()
task_configs = task_config.repeat(1,len(rewards)).view(-1,3)
log_softmax_actions = actor_network(states_var)
vs = meta_value_network(torch.cat((states_var,task_configs.detach()),1)).detach()
# calculate qs
qs = Variable(torch.Tensor(discount_reward(rewards,0.99,final_r))).cuda()
advantages = qs - vs
actor_network_loss = - torch.mean(torch.sum(log_softmax_actions*actions_var,1)* advantages) #+ entropy #+ actor_criterion(actor_y_samples,target_y)
actor_network_loss.backward()
torch.nn.utils.clip_grad_norm(actor_network.parameters(),0.5)
actor_network_optim.step()
pre_states = states
pre_actions = actions
pre_rewards = rewards
# testing
if (step + 1) % 10 == 0:
# testing
result = 0
test_task = CartPoleEnv(length = task.length)
for test_epi in range(10):
state = test_task.reset()
for test_step in range(200):
softmax_action = torch.exp(actor_network(Variable(torch.Tensor([state])).cuda()))
#print(softmax_action.data)
action = np.argmax(softmax_action.cpu().data.numpy()[0])
next_state,reward,done,_ = test_task.step(action)
result += reward
state = next_state
if done:
break
aver_result = result/10.0
test_results.append(aver_result)
train_games.append(task.episodes)
print("task length:",task_length,"episode:",episode,"step:",step+1,"result:",aver_result)
data_i_episode["test_results"] = test_results
data_i_episode["train_games"] = train_games
data_i["results"] = data_i_episode
datas.append(data_i)
save_to_json('mvn_cartpole_test_100.json', datas)
for i in range(episodes):
print("episode=", i)
obs = env.reset()
s = discretize_state(obs, s_bounds, n_s)
finished = False
time_step = 0
#dlugie epizody przerywamy po 200 krokach
while not finished and not time_step == 200:
#polityka stochastyczna - prawdopodobienstwo 0.5 dla obu akcji (w lewo, w prawo)
action = np.random.randint(0, 2)
#DO UZUPEŁNIENIA
obs, reward, done, info = env.step(action)
s_ = discretize_state(obs, s_bounds, n_s)
V[s] = V[s] + alpha * (reward + gamma * V[s_] - V[s])
s = s_ #s = s'
if done == True:
finished = True
break
time_step += 1
print(V)
# In[ ]:
class CuteLearning():
def __init__(self):
self.plot_data = PlotData()
self.env = CartPoleEnv()
self.main_net = DQN()
self.target_net = deepcopy(self.main_net)
self.epsilon = config.epsilon
self.eps_decay = 0.995
self.visu = False
self.visu_update = False#300
self.visu_window = 5
self.memory = Memory(memory_size = 30)
self.batch_size = 5
def reward_optimisation(self, state, end):
reward = 0 if end else 1
return reward
def choose_action(self, q_values):
if (random.random() > self.epsilon):
return(np.argmax(q_values))
else:
return random.randint(0,1)
def make_batch(self):
batch = self.memory.get_batch(self.batch_size)
states = []
targets = []
for s, a, r, ns, done in batch:
states.append(s)
q_values = self.target_net.predict(s).tolist()
if done:
q_values[a] = r
else:
q_values_next = self.target_net.predict(ns)
q_values[a] = r + net_config.gamma * torch.max(q_values_next).item()
targets.append(q_values)
return states, targets
def updato(self):
states, targets = self.make_batch()
self.main_net.update(states, targets)
def upodato(self, state, reward, next_state, done):
def learn(self, episodes = 10000, replay = False):
episode = 0
tmp = self.epsilon
while (episode < episodes):
done = False
turn = 0
state = self.env.reset()
self.eps_decay = self.epsilon * self.eps_decay
while (done == False):
q_values = self.main_net.model(torch.Tensor(state)).tolist()
action = self.choose_action(q_values)
new_state, reward, done, _ = self.env.step(action)
self.memory.add_data((state, action, reward, new_state, done))
state = new_state
turn += 1
self.updato()
print("turn:", turn)
episode += 1
if (episode % net_config.n_update == 0):
self.target_net = deepcopy(self.main_net)
self.epsilon = tmp
def save(self):
pass
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
Cutie = CuteLearning()
Cutie.learn()
class CuteLearning():
def __init__(self):
self.plot_data = PlotData()
self.cart = CartPoleEnv()
self.cart.reset()
self.predi_net = DQN()
self.updat_net = deepcopy(self.predi_net)
self.turn = 0
self.epidode = 0
self.epsilon = config.epsilon
self.eps_decay = 0.99
self.visu = False
self.visu_update = False #300
self.visu_window = 5
self.consecutive_wins = 0
self.best_consecutive_wins = 0
self.last_save = 0
self.memory = []
def reward_optimisation(self, state, end):
reward = -25 if end else 1
if reward == 1:
# Angle reward modification
angle_r = 0.418 / 2
reward += (((abs(angle_r - abs(state[2])) / angle_r) * 2) - 1) * 2
# Position reward modification
pos_r = 0.418 / 2
reward += (((abs(pos_r - abs(state[0])) / pos_r) * 2) - 1) * 2
return reward
def learn(self):
self.episode = 0
n = 0
while self.episode < 10:
self.turn = 0
end = False
states = []
targets = []
while not end:
# 1. Init
state = self.cart.state
# 2. Choose action
q_values = self.predi_net.predict(state).tolist()
a = choose_action_net(q_values, self.epsilon)
# 3. Perform action
next_state, _, end, _ = self.cart.step(a)
# 4. Measure reward
reward = self.reward_optimisation(next_state, end)
q_values_next = self.predi_net.predict(next_state)
# 5. Calcul Q-Values
q_values[a] = reward + net_config.gamma * \
torch.max(q_values_next).item()
self.turn += 1
self.memory.append((state, a, next_state, reward, end))
# self.updat_net.update(state, q_values)
states.append(state)
targets.append(q_values)
if (self.turn % 20 and self.turn) or end:
self.updat_net.update(states, targets)
states = []
targets = []
if self.turn >= 500:
end = True
if self.visu:
self.cart.render()
self.episode += 1
self.replay(20)
if self.episode % net_config.n_update == 0 and self.episode:
print("Update")
self.predi_net.model.load_state_dict(
self.updat_net.model.state_dict())
self.end()
n += 1
self.save()
self.cart.close()
self.plot_data.clear()
def replay(self, size):
if size > len(self.memory):
size = len(self.memory)
data = random.sample(self.memory, size)
states = []
targets = []
for state, action, next_state, reward, done in data:
q_values = self.predi_net.predict(state)
if done:
q_values[action] = reward
else:
# The only difference between the simple replay is in this line
# It ensures that next q values are predicted with the target network.
q_values_next = self.predi_net.predict(next_state)
q_values[action] = reward + net_config.gamma * torch.max(
q_values_next).item()
states.append(state)
targets.append(q_values)
self.updat_net.update(state, q_values)
def end(self):
self.plot_data.new_data(self.turn)
if self.turn > 195:
self.consecutive_wins += 1
if self.best_consecutive_wins < self.consecutive_wins:
self.best_consecutive_wins = self.consecutive_wins
if self.consecutive_wins > 200:
self.save()
print(("WIN IN " + str(self.episode) + " EPISODES\n") * 100)
else:
self.consecutive_wins = 0
if self.last_save * 1.2 < self.best_consecutive_wins and 50 <= self.best_consecutive_wins:
self.save()
self.last_save = self.best_consecutive_wins
print("Episode: ", self.episode, "\tTurn:", self.turn, "\tEpsilon:",
self.epsilon, "\tWins: ", "{:3}".format(self.consecutive_wins),
"/", self.best_consecutive_wins)
self.turn = 0
self.cart.reset()
if self.episode % config.graph_update == 0 and self.episode != 0:
self.plot_data.graph()
if self.visu_update:
if self.episode % self.visu_update == 0:
self.visu = True
if self.episode % self.visu_update == self.visu_window:
self.visu = False
self.cart.close()
self.epsilon = max(self.epsilon * self.eps_decay, 0.01)
def save(self):
pass
import gym
from cartpole import CartPoleEnv
env = CartPoleEnv()
observation = env.reset()
total_reward = 0
print(env.observation_space)
print(env.observation_space.high)
print(env.observation_space.low)
for t in range(200):
env.render()
action = env.action_space.sample()
observation, reward, done, info = env.step(action)
total_reward += reward
if done:
break
env.close()
print(total_reward)
agent = RandomAgent(env.action_space)
episode_count = 100
reward = 0
done = False
action_record = []
state_record = []
for i in range(episode_count):
ob = env.reset()
#action = agent.act(ob, reward, done)
traj_action = []
traj_state = []
while True:
action = agent.act(ob, reward, done)
ob, reward, done, _ = env.step(action)
traj_action.append(action)
traj_state.append(ob)
print(done)
#env.render()
if done:
break
action_record.append(traj_action)
state_record.append(traj_state)
action_record = np.asarray(action_record)
state_record = np.asarray(state_record)
record = {'actions': action_record, 'states':state_record}
record_file = open('record.pickle', 'wb')
pickle.dump(record, record_file)
# Note there's no env.render() here. But the environment still can open window and