def evaluate(idx, target_value_network, lock, counter):
port = 6000 + idx
seed = 123 + idx * 10
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed)
hfoEnv.connectToServer()
thread_counter = 0
target_value_network.load_state_dict(torch.load('checkpoint/checkpoint'))
num_episode = 0
goal_list = []
while True:
epsilon = 0.
print('goal_list\n\n\n', goal_list)
done = False
curState = hfoEnv.reset()
timestep = 0
while not done and timestep < 500:
action = epsilon_greedy(curState, epsilon, target_value_network)
nextState, reward, done, status, info = hfoEnv.step(
hfoEnv.possibleActions[action])
curState = nextState
with lock:
counter.value += 1
thread_counter += 1
timestep += 1
if status == GOAL:
goal_list.append(num_episode)
num_episode += 1
def create_greedy_worker(networks, counter, config):
logger = Logger("output_{0}_greedy.out".format(config['experiment']))
environment = HFOEnv(port=6321, seed=86868686, numOpponents=1)
environment.connectToServer()
w_args = (100000, networks["learning"], environment, Policy(logger=logger),
logger, counter)
return mp.Process(target=policy_worker.run, args=w_args)
def train(idx, args, value_network, target_value_network, optimizer, lock,
counter):
port = args.port
seed = args.seed
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed)
hfoEnv.connectToServer()
eps = args.epsilon
I_target = args.I_target
I_update = args.I_update
discountFactor = args.discountFactor
learn_step_counter = 0
loss_func = nn.MSELoss()
threads = args.numprocesses
# This runs a random agent
s = hfoEnv.reset()
while learn_step_counter < counter.value / threads:
lock.acquire()
counter.value += 1
lock.release()
learn_step_counter += 1
# take action
action = choose_action(s, value_network, eps)
# nextState, reward, done, status, info based on the action
s_, r, done, status, info = hfoEnv.step(action)
act_index = hfoEnv.possibleActions.index(action)
q_eval = computePrediction(s, act_index, value_network)
q_target = computeTargets(r, s_, discountFactor, done,
target_value_network)
loss = loss_func(q_eval, q_target)
loss.backward()
if learn_step_counter % I_update or done:
#lock.acquire()
optimizer.step()
optimizer.zero_grad()
optimizer.share_memory()
#lock.release()
# target parameter update
if counter.value % I_target == 0:
lock.acquire()
target_value_network.load_state_dict(value_network.state_dict())
lock.release()
if counter.value == 1e6:
saveModelNetwork(target_value_network, os.getcwd())
if done:
s = hfoEnv.reset()
s = s_
def train(idx, args, value_network, target_value_network, optimizer, lock,
counter):
port = 6000 + idx
seed = 123 + idx * 10
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed)
hfoEnv.connectToServer()
thread_counter = 0
criterion = nn.MSELoss()
optimizer.zero_grad()
target_value_network.load_state_dict(torch.load('params/params_last'))
num_episode = 0
while True:
epsilon = args.epsilon * (
(1 - 1 / (1 + np.exp(-thread_counter / 2000))) * 2 * 0.9 + 0.1)
done = False
curState = hfoEnv.reset()
timestep = 0
while not done and timestep < 500:
action = epsilon_greedy(curState, epsilon, value_network)
nextState, reward, done, status, info = hfoEnv.step(
hfoEnv.possibleActions[action])
pred_value = computePrediction(curState, action, value_network)
target_value = computeTargets(reward, nextState,
args.discountFactor, done,
target_value_network)
loss = criterion(pred_value, target_value)
loss.backward()
curState = nextState
with lock:
counter.value += 1
thread_counter += 1
timestep += 1
if counter.value % args.iterate_target == 0:
target_value_network.load_state_dict(
torch.load('params/params_last'))
if thread_counter % args.iterate_async == 0 or done:
optimizer.step()
optimizer.zero_grad()
saveModelNetwork(value_network, 'params/params_last')
if counter.value % 1000000 == 0:
saveModelNetwork(
value_network,
'params/params_{0:d}'.format(counter.value // 1000000))
num_episode += 1
def train(idx, args, value_network, target_value_network, optimizer, lock, counter):
port = 3020+idx*10
seed = 12+idx*10
counter_v = 0
hfoEnv = HFOEnv(numTeammates = 0, numOpponents =1, port = port, seed = seed)
hfoEnv.connectToServer()
loss = nn.MSELoss()
for episodeNumber in range(1,args.epochs+1):
observation = hfoEnv.reset()
counter.value += 1
counter_v += 1
for timestep in range(args.timesteps):
#observation_t = torch.Tensor(observation)
#action = greedy_action(observation_t,value_network,args)
#print('!!!!!!!!!!!!!!',action)
#act = hfoEnv.possibleActions[action]
action = random.randint(0,3)
act = hfoEnv.possibleActions[action]
newObservation, reward, done, status, info = hfoEnv.step(act)
#print(newObservation,reward,done,status,info)
optimizer.zero_grad()
lock.acquire()
observation_t = torch.Tensor(observation)
newObservation_t = torch.Tensor(newObservation)
output = computePrediction(observation_t,action,value_network)
target = computeTargets(reward,newObservation_t,args.discountFactor,done,target_value_network)
lock.release()
out = loss(output,target)
out.backward()
if counter_v% args.target_interval == 0:
target_value_network.load_state_dict(value_network.state_dict())
if counter.value % args.predict_interval == 0 or done:
optimizer.step()
optimizer.zero_grad()
strDirectory = "value_network"+str(idx)+".pth"
saveModelNetwork(value_network,strDirectory)
strDirectory_targert = "targetNetwork"+str(idx)+".pth"
saveModelNetwork(target_value_network,strDirectory_targert)
observation = newObservation
if done:
break
if counter.value >= args.tmax:
break
#!/usr/bin/env python3
# encoding utf-8
import os
import policy_worker
from Environment import HFOEnv
import network_factory
import torch
from Policy import RandomPolicy
os.environ['OMP_NUM_THREADS'] = '1'
# Use this script to handle arguments and
# initialize important components of your experiment.
# These might include important parameters for your experiment,
# your models, torch's multiprocessing methods, etc.
if __name__ == "__main__":
rnd_seed = 11111 + 111
environment = HFOEnv(port=6011, seed=rnd_seed, numOpponents=1)
environment.connectToServer()
policy_worker.run(num_episodes=1000,
value_network={},
environment=environment,
policy=RandomPolicy())
environment.quitGame()
def train(idx, args, value_network, target_value_network, optimizer, lock, counter,
port, seed, I_tar, I_async, name=None):
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed, headless=True)
hfoEnv.connectToServer()
if name == None:
name = str(time())
#logger = Logger("tb/" + name, flush_secs=5)
target_value_network.train()
num_episodes = args.episodes
gamma = 0.99
windows = [10, 500]
goal_buffer = [0]*max(windows)
max_epsilon = 0.99
min_epsilon = 0.01
total = 200
epsilon_fn = lambda current : max_epsilon - current*(max_epsilon - min_epsilon)/total if current < total else min_epsilon
max_lr = 0.9
min_lr = 0.1
total = 600
lr_fn = lambda current: max_lr - current*(max_lr - min_lr)/total if current < total else min_lr
loss_func = nn.MSELoss()
t = 0
episodeNumber = 0
episodeReward = 0
episodeSteps = 0
state1 = hfoEnv.reset()
while episodeNumber <= num_episodes:
if (counter.value+1) % 1e3 == 0:
print("##################################################")
print('t:',t)
print('counter:',counter.value)
print("##################################################")
if (counter.value+1) % 1e6 == 0:
saveModelNetwork(value_network,"trained_models/params"+str(int((counter.value+1) // 1e6)))
#train_epoch(epoch, args, model, device, train_loader, optimizer)
epsilon = epsilon_fn(episodeNumber)
lr = lr_fn(episodeNumber)
if args.eval or np.random.random() >= epsilon:
qs = value_network(state1)
action = torch.argmax(qs)
else:
action = random.randint(0,len(hfoEnv.possibleActions)-1)
a1 = hfoEnv.possibleActions[action]
state2, reward, done, status, info = hfoEnv.step(a1)
episodeReward += reward
y = computeTargets(reward, state2, gamma, done, target_value_network)
prediction = computePrediction(state1,action,value_network)
Y = torch.zeros(4)
Y[action] = y
Prediction = torch.zeros(4)
Prediction[action] = prediction
loss = loss_func(Y,Prediction)
loss.backward()
state1 = state2
t += 1
episodeSteps += 1
with lock:
counter.value = counter.value + 1
if done:
if status == GOAL:
goal_buffer.append(1)
else:
goal_buffer.append(0)
#logger.log_value('episode/reward',episodeReward, episodeNumber)
#logger.log_value('episode/length',episodeSteps, episodeNumber)
#logger.log_value('hyperparameters/epsilon', epsilon, episodeNumber)
#logger.log_value('hyperparameters/lr', lr, episodeNumber)
#for window in windows:
# logger.log_value(learning_str + "goals/%i" % window,
# np.sum(goal_buffer[-window:]),
# episodeNumber)
episodeNumber += 1
episodeReward = 0.0
episodeSteps = 0
state1 = hfoEnv.reset()
if t % I_async == 0 or done or episodeNumber == num_episodes:
# Async update of value_network using gradients
with lock:
# Add grads to value_network
for param, shared_param in zip(
value_network.parameters(),
target_value_network.parameters()):
shared_param._grad = param.grad
#value_network._grad = target_value_network.grad
# Take a step
optimizer.step(lr=lr)
# Clean gradients
optimizer.zero_grad()
target_value_network.zero_grad()
#if counter.value % I_tar == 0 or episodeNumber == num_episodes:
if t % I_tar == 0 or episodeNumber == num_episodes or done:
# Update target network
target_value_network.zero_grad()
target_value_network.load_state_dict(value_network.state_dict())
hfoEnv.reset()
saveModelNetwork(value_network,"trained_models/params_last")
# Finishing training and showing stats
hfoEnv.quitGame()
def train(idx, args, value_network, target_value_network, optimizer, lock,
counter):
port = 8000 + 50 * idx
env = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=123)
env.connectToServer()
do_opt_step = False
do_hard_copy = False
total_reward = 0
save_counter = 1
for episode in range(args.numEpisodes):
timestep = 0
total_reward = 0
obs = env.reset()
done = False
save_model = False
do_opt_step = False
do_hard_copy = False
while not done and timestep < 500:
# obs to tensor
obs_tensor = torch.Tensor(obs).unsqueeze(0)
# choose action
act = chooseAction(obs_tensor, value_network, episode, idx)
# execute action
act_str = env.possibleActions[act]
next_obs, rewards, done, status, info = env.step(act_str)
# update total reward
total_reward += rewards
# reward to tensor
reward_tensor = torch.Tensor([rewards])
# next obs to tensor
next_obs_tensor = torch.Tensor(next_obs).unsqueeze(0)
# update counters and flags
timestep += 1
with lock:
counter.value = counter.value + 1
if timestep % args.aSyncFreq == 0 or done:
do_opt_step = True
if counter.value % args.copy_freq == 0:
do_hard_copy = True
if counter.value % 1e6 == 0:
save_model = True
current_count = counter.value
#forward pass for our networks
predicted_vals = computePrediction(obs_tensor, act, value_network)
target_vals = computeTargets(reward_tensor, next_obs_tensor,
args.gamma, done,
target_value_network)
# loss function calculation
loss_function = nn.MSELoss()
err = loss_function(predicted_vals, target_vals)
# accumulate gradients
err.backward()
# update optimizer
if do_opt_step:
with lock:
optimizer.step()
optimizer.zero_grad()
do_opt_step = False
#update global network
if do_hard_copy:
with lock:
hard_copy(target_value_network, value_network)
do_hard_copy = False
# update current state
obs = next_obs
if save_model:
#save model
saveModelNetwork(value_network, 'params_' + str(save_counter))
save_counter += 1
#change learning rate
change_lr(current_count, optimizer)
saveModelNetwork(value_network, 'params_latest')
def train(idx, val_network, target_value_network, optimizer, lock, counter,timesteps_per_process,results_dir):
# This runs a random agent
episodeNumber = 0
discountFactor = 0.99
f = open(os.path.join(results_dir, 'worker_%d.out'%idx), 'w')
#Each worker writes to a file to check if my network runs correctly
columns = '{0:<10} {1:<8} {2:<10} {3:<12} {4:<20} {5:<15} {6:<15}\n'
f.write(columns.format('Episode','Status','Steps','Total steps','Avg steps to goal','Total Goals','Counter'))
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=6000+(idx+5)*10, seed=idx)
hfoEnv.connectToServer()
asyc_update = 10
copy_freq = 10000
total_goals = 0
total_steps = 0
totalEpisodes = timesteps_per_process // 500
save_model = False
copy_model = False
for episodeNumber in range(totalEpisodes):
done = False
epsilon = epsilon_annealing(idx,episodeNumber,totalEpisodes)
state = hfoEnv.reset()
timesteps = 0
while timesteps<500:
lock.acquire()
counter.value += 1
locked_counter = counter.value
if (locked_counter % 1e6) == 0:
save_model = True
if (locked_counter % copy_freq) == 0:
copy_model = True
lock.release()
timesteps+=1
obs_tensor = torch.Tensor(state).unsqueeze(0)
Q, act = compute_val(val_network, obs_tensor,idx,epsilon)
act = hfoEnv.possibleActions[act]
newObservation, reward, done, status, info = hfoEnv.step(act)
Q_target = computeTargets(reward,torch.Tensor(newObservation).unsqueeze(0),discountFactor,done,target_value_network)
loss_function = nn.MSELoss()
loss = loss_function(Q_target,Q)
loss.backward()
if timesteps % asyc_update ==0 or done:
with lock:
optimizer.step()
optimizer.zero_grad()
if copy_model:
hard_copy(target_value_network, val_network)
copy_model = False
if save_model:
updating_learning_rate(optimizer,locked_counter)
saveModelNetwork(target_value_network,os.path.join(results_dir, 'params_%d' % int(locked_counter / 1e6)))
save_model = False
state = newObservation
if done:
break
if status==1:
total_steps += timesteps
total_goals += 1
else:
total_steps += 500
f.write(columns.format(episodeNumber, status, timesteps, total_steps, '%.1f'%(total_steps/(episodeNumber+1)), total_goals, locked_counter))
f.flush()
def runTrain(idx, args, valueNetwork, targetNetwork, optimizer, lock, counter):
if args.use_gpu and torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
port = 6006 + idx * 9
seed = 2019 + idx * 46
hfoEnv = HFOEnv(args.reward_opt,
numTeammates=0,
numOpponents=1,
port=port,
seed=seed)
hfoEnv.connectToServer()
episodeNumber = 0
numTakenActions = 0
numTakenActionCKPT = 0
discountFactor = args.discountFactor
totalWorkLoad = args.t_max // args.n_jobs
steps_to_ball = [
] # number of steps spent to approach the ball in each episode
steps_in_episode = [] # number of steps in each episode
status_lst = []
log = {
'steps_to_ball': steps_to_ball,
'steps_in_episode': steps_in_episode,
'status_lst': status_lst
}
step_to_ball = None
firstRecord = True
optimizer.zero_grad()
state = torch.tensor(hfoEnv.reset()).to(device) # get initial state
while True:
# take action based on eps-greedy policy
action = select_action(state, valueNetwork, episodeNumber,
numTakenActions, args) # action -> int
act = hfoEnv.possibleActions[action]
newObservation, reward, done, status, info = hfoEnv.step(act)
if info['kickable'] and firstRecord:
step_to_ball = numTakenActions - numTakenActionCKPT
firstRecord = False
next_state = torch.tensor(
hfoEnv.preprocessState(newObservation)).to(device)
target = computeTargets(reward, next_state, discountFactor, done,
targetNetwork, device) # target -> tensor
pred = computePrediction(state, action, valueNetwork,
device) # pred -> tensor
loss = F.mse_loss(pred, target) # compute loss
loss.backward() # accumulate loss
lock.acquire()
counter.value += 1
counterValue = counter.value
lock.release()
numTakenActions += 1
if done:
firstRecord = True
status_lst.append(status)
steps_in_episode.append(numTakenActions - numTakenActionCKPT)
if step_to_ball is None:
steps_to_ball.append(numTakenActions - numTakenActionCKPT)
else:
steps_to_ball.append(step_to_ball)
step_to_ball = None
episodeNumber += 1
numTakenActionCKPT = numTakenActions
# hfoEnv alreay call 'preprocessState' in 'reset'
state = torch.tensor(hfoEnv.reset()).to(device)
if done or numTakenActions % args.i_async_update == 0:
optimizer.step() # apply grads
optimizer.zero_grad() # clear all cached grad
if counterValue % args.i_target == 0:
targetNetwork.load_state_dict(
valueNetwork.state_dict()) # update target network
if counterValue % args.ckpt_interval == 0:
ckpt_path = os.path.join(args.log_dir, 'ckpt')
if not os.path.exists(ckpt_path):
os.mkdir(ckpt_path)
filename = os.path.join(
ckpt_path, 'params_%d' % (counterValue / args.ckpt_interval))
lock.acquire()
saveModelNetwork(valueNetwork, filename)
lock.release()
if numTakenActions > totalWorkLoad:
filename = os.path.join(args.log_dir, 'log_worker_%d.pkl' % idx)
saveLog(log, filename)
return
def runEval(args, valueNetwork):
port = 6050
seed = 2019
if args.use_gpu and torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
hfoEnv = HFOEnv(args.reward_opt,
numTeammates=0,
numOpponents=1,
port=port,
seed=seed)
hfoEnv.connectToServer()
episodeNumber = 0
numTakenActions = 0
numTakenActionCKPT = 0
state = torch.tensor(hfoEnv.reset()).to(device) # get initial state
steps_to_ball = []
steps_in_episode = []
status_lst = []
log = {
'steps_to_ball': steps_to_ball,
'steps_in_episode': steps_in_episode,
'status_lst': status_lst
}
step_to_ball = None
firstRecord = True
while numTakenActions < args.t_max:
# take action based on eps-greedy policy
action = select_action(state, valueNetwork, episodeNumber,
numTakenActions, args) # action -> int
act = hfoEnv.possibleActions[action]
newObservation, reward, done, status, info = hfoEnv.step(act)
if info['kickable'] and firstRecord:
step_to_ball = numTakenActions - numTakenActionCKPT
firstRecord = False
# all updates on the parameters shall acqure lock
numTakenActions += 1
if done:
firstRecord = True
status_lst.append(status)
steps_in_episode.append(numTakenActions - numTakenActionCKPT)
if step_to_ball is None:
steps_to_ball.append(numTakenActions - numTakenActionCKPT)
else:
steps_to_ball.append(step_to_ball)
step_to_ball = None
episodeNumber += 1
numTakenActionCKPT = numTakenActions
# hfoEnv alreay call 'preprocessState' in 'reset'
state = torch.tensor(hfoEnv.reset()).to(device)
filename = os.path.join(args.log_dir, 'log_eval.pkl')
saveLog(log, filename)
(total_goals, num_episodes, total_steps, steps_per_goal))
f.flush()
f.close()
value_network = ValueNetwork()
idx = 12
port = 8000 + idx * 20
seed = idx * 2
directories = [x[0] for x in os.walk('.') if 'results' in x[0]]
print('starting HFO')
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed)
hfoEnv.connectToServer()
for i in range(4, 8):
fn = 'params_' + str(i)
print('Loading from ', fn)
value_network.load_state_dict(torch.load(fn))
evaluate(value_network, hfoEnv, 500, 'results_' + str(i))
import os
directories = [x[0] for x in os.walk('.') if 'results' in x[0]]
for d in directories:
fn = os.path.join(d, 'evaluation_best.out')
f = open(fn, 'r')
print(d)
for ln in f:
def train(idx, args, learning_network, target_network, optimizer, lock,
counter):
# init port & seed for the thread based on id val
port = 8100 + 10 * idx # init
seed = idx * 113 + 923
torch.manual_seed(seed)
worker_network = ValueNetwork(15, 4)
worker_network.load_state_dict(learning_network.state_dict())
# change init
# init env
hfo_env = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed)
hfo_env.connectToServer()
episode_num = 0
eps = random.sample(epsilon_list, 1)[0]
worker_timestep = 0
mse_loss = nn.MSELoss()
max_worker_steps = args.max_steps / args.num_processes
can_continue = True
goal = 0
to_goal = []
while can_continue:
# run episode
obs_tensor = hfo_env.reset()
done = False
loss = 0
reward_ep = 0
ep_steps = 0
upd_steps = 0
while not done:
# select action based on greedy policy
action_idx = select_action(obs_tensor, worker_network,
worker_timestep, max_worker_steps, args,
eps)
action = hfo_env.possibleActions[action_idx]
# observe next
next_obs_tensor, reward, done, status, info = hfo_env.step(action)
y = computeTargets(reward, next_obs_tensor, args.discount, done,
target_network)
# compute predictions again for the best action. Here we change params
q_next = computePrediction(obs_tensor, action_idx, worker_network)
# put new state
obs_tensor = next_obs_tensor
# update episode stats
loss += mse_loss(y, q_next)
reward_ep += reward
upd_steps += 1
ep_steps += 1
worker_timestep += 1
if status == 1:
goal += 1
to_goal.append(ep_steps)
with lock:
counter.value += 1
if counter.value % args.checkpoint_time == 0:
saveModelNetwork(
learning_network,
args.checkpoint_dir + '_{}'.format(counter.value))
# if terminal or time to update network
if done or worker_timestep % args.val_net_update_freq == 0:
worker_network.zero_grad()
optimizer.zero_grad()
# take mean loss
loss /= upd_steps
loss.backward()
sync_grad(learning_network, worker_network)
optimizer.step()
worker_network.load_state_dict(learning_network.state_dict())
loss = 0
upd_steps = 0
# perform update of target network
if counter.value % args.tgt_net_update_freq == 0:
target_network.load_state_dict(learning_network.state_dict())
if counter.value % args.checkpoint_time == 0:
saveModelNetwork(
learning_network,
args.checkpoint_dir + '_{}'.format(counter.value))
episode_num += 1
# if time is exceeded -> break the loop
can_continue = counter.value <= args.max_steps\
and worker_timestep <= max_worker_steps\
and status !=SERVER_DOWN \
and episode_num <= args.num_episodes # lets run just 8K episodes
# finish the game
hfo_env.quitGame()
# save the network it stopped with
saveModelNetwork(learning_network,
args.checkpoint_dir + '_{}_final'.format(counter.value))
def train(idx,
port,
target_network,
value_network,
lock,
counter,
num_episodes=16000,
name=""):
print("Starting a worker {}".format(port))
# port = 2207
seed = 2207
hfoEnv = HFOEnv(numTeammates=0, numOpponents=1, port=port, seed=seed)
hfoEnv.connectToServer()
episodeNumber = 0
epsilon = 1
discountFactor = 0.99
I_async_update = 5
I_target = 10000
goals = 0
paramSaves = 0
lastSaved = 0
hard_update(target_network, value_network)
optimizer = optim.Adam(value_network.parameters(), lr=1e-5)
optimizer.zero_grad()
t = 0 # local timestep counter
if idx == 0: # first thread keeps track of stats
stats = []
# run for certain number of timesteps
while t < num_episodes * 500:
timesteps_to_goal = 0 # for measuring performance
total_reward = 0 # accumulated reward (without discounting)
status = 0
observation = hfoEnv.reset()
# print(observation.shape)
# linearly decrease epsilon
epsilon = max(0.0, (22000 - episodeNumber) / 22000)
while status == 0:
# EPSILON GREEDY - TAKE AN ACTION
if np.random.rand() < epsilon:
# choose a random action
action = np.random.choice(range(len(hfoEnv.possibleActions)))
else:
# choose greedy action
lock.acquire()
values = value_network(
torch.Tensor(observation)).detach().numpy()
action = np.argmax(values)
lock.release()
newObservation, reward, done, status, info = hfoEnv.step(
hfoEnv.possibleActions[action])
total_reward += reward
# keep track of goals scored
if reward >= 50.0:
goals += 1
# COMPUTE TARGET VALUE
lock.acquire()
target_value = computeTargets(reward, [newObservation],
discountFactor, done, target_network)
prediction = computePrediction([observation], action,
value_network)
loss = 0.5 * (prediction - target_value.detach())**2
# accummulate gradient
loss.backward()
lock.release()
observation = newObservation
# update local counter t
t += 1
timesteps_to_goal += 1
# update global counter T
lock.acquire()
counter.value += 1
# update target network
if counter.value % I_target == 0:
hard_update(target_network, value_network)
# only the first worker saves the model (every 1 mil)
if idx == 0 and counter.value >= 1000000 + lastSaved:
lastSaved = counter.value
print("saving model")
paramSaves += 1
path = "{}_params_{}".format(name, paramSaves)
saveModelNetwork(value_network, path)
# update value network and zero gradients
if t % I_async_update == 0 or done:
print("Doing async update")
optimizer.step()
optimizer.zero_grad()
lock.release()
if done:
if idx == 0:
timesteps_to_goal = timesteps_to_goal if status == 1 else 500
stats.append(timesteps_to_goal)
mean = np.mean(stats) # mean ep length
# output things to a csv for monitoring
print("{}, {}, {}, {}, {}, {}, {}, {}".format(
episodeNumber, t, mean, goals, epsilon,
timesteps_to_goal, status, total_reward),
file=open("{}experiment.csv".format(name), "a"))
episodeNumber += 1
def create_environment(idx):
port = 6000 + idx * 4
rnd_seed = 11111 * idx + 111
environment = HFOEnv(port=port, seed=rnd_seed, numOpponents=1)
environment.connectToServer()
return environment