def runEpisode(step_limit):
global which_episode
global total_win
global total_draw
global total_lose
global ns_epoch
global pcts_win
global pcts_win_or_draw
global pcts_lose
which_episode += 1
# ゲーム1回 開始
terminal = RLGlue.RL_episode(step_limit)
# 勝負がつくまでのステップ数と報酬を取得
total_steps = RLGlue.RL_num_steps()
total_reward = RLGlue.RL_return()
# 今回の結果を表示
r_win = 1.0
r_draw = -0.5
r_lose = -1.0
if total_reward == r_win:
total_win += 1
elif total_reward == r_draw:
total_draw += 1
elif total_reward == r_lose:
total_lose += 1
print("Episode " + str(which_episode) + "\t " + str(total_steps) +
" steps \t" + str(total_reward) + " total reward\t " +
str(terminal) + " natural end")
# 100回毎に勝敗を集計
record_interval = 100
if which_episode % record_interval == 0:
line = 'Episode: {}, {} wins, {} draws, {} loses'.format(
which_episode, total_win, total_draw, total_lose)
print(
'---------------------------------------------------------------')
print(line)
print(
'---------------------------------------------------------------')
# 集計結果をファイルに出力
with open('result.txt', 'a') as f:
f.writelines(line + '\n')
ns_epoch.append(which_episode)
pcts_win.append(float(total_win) / record_interval * 100)
pcts_win_or_draw.append(
float(total_win + total_draw) / record_interval * 100)
pcts_lose.append(float(total_win) / record_interval * 100)
total_win = 0
total_draw = 0
total_lose = 0
def main():
"""
Run the desired number of training epochs, a testing epoch
is conducted after each training epoch.
"""
parser = argparse.ArgumentParser(description='Neural rl experiment.')
parser.add_argument('--num_epochs',
type=int,
default=100,
help='Number of training epochs')
parser.add_argument('--epoch_length',
type=int,
default=50000,
help='Number of steps per epoch')
parser.add_argument('--test_length',
type=int,
default=10000,
help='Number of steps per test')
args = parser.parse_args()
RLGlue.RL_init()
for epoch in range(1, args.num_epochs + 1):
RLGlue.RL_agent_message("training")
run_epoch(epoch, args.epoch_length, "training")
RLGlue.RL_agent_message("start_testing")
run_epoch(epoch, args.test_length, "testing")
RLGlue.RL_agent_message("finish_testing " + str(epoch))
def run_episode(self):
""" Run a single episode """
# Update epsilon
'''
phase_len = self.episodes / 3
if self.episode_number == phase_len * 2:
# Start low phase
RLGlue.RL_agent_message('set epsilon %f' % (self.epsilon_low))
elif self.episode_number >= phase_len and self.episode_number < phase_len * 2:
# In decr phase
epsilon = float(RLGlue.RL_agent_message('get epsilon'))
epsilon += (self.epsilon_decr - self.epsilon_high) / phase_len
RLGlue.RL_agent_message('set epsilon %f' % (epsilon))
elif self.episode_number == 0:
# Start high phase
RLGlue.RL_agent_message('set epsilon %f' % (self.epsilon_high))
'''
terminal = RLGlue.RL_episode(0) # 0 - run until terminal
steps = RLGlue.RL_num_steps()
reward = RLGlue.RL_return()
#print "\nEpisode %d\t %d steps\t reward: %d" % (self.episode_number, steps, reward)
#print "Episode "+str(episode_number)+"\t "+str(totalSteps)+ " steps \t" + str(totalReward) + " total reward\t " + str(terminal) + " natural end"
self.returns[self.episode_number] = (
reward + self.returns[self.episode_number] *
(self.instance - 1)) / self.instance
self.steps[self.episode_number] = (steps +
self.steps[self.episode_number] *
(self.instance - 1)) / self.instance
self.episode_number += 1
def run_epoch(self, epoch, num_steps, prefix, collect_reward=False):
""" Run one 'epoch' of training or testing, where an epoch is defined
by the number of steps executed. Prints a progress report after
every trial
Arguments:
num_steps - steps per epoch
prefix - string to print ('training' or 'testing')
"""
steps_left = num_steps
if prefix == "training" or not collect_reward:
while steps_left > 0:
print prefix + " epoch: ", epoch, "steps_left: ", steps_left
sys.stdout.flush()
terminal = RLGlue.RL_episode(steps_left)
if not terminal:
RLGlue.RL_agent_message("episode_end")
steps_left -= RLGlue.RL_num_steps()
elif prefix == "testing":
total_reward = 0
episode_counter = 0
terminal = False
while steps_left > 0:
if terminal:
print prefix + " epoch: ", epoch, "steps_left: ", steps_left
sys.stdout.flush()
roat = RLGlue.RL_step()
reward = roat.r
terminal = roat.terminal
total_reward += reward
episode_counter += terminal
steps_left -= 1
return total_reward, episode_counter
def runEpisode(is_learning_episode):
global whichEpisode, learningEpisode
RLGlue.RL_episode(10000)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
whichEpisode += 1
if is_learning_episode:
learningEpisode += 1
#print "Episode " + str(learningEpisode) + "/" + str(whichEpisode) + "\t " + str(totalSteps) + " steps \t" + str(totalReward) + " total reward\t "
print "Episode %d/%d\t %d steps \t %.1f total reward\t" % (
learningEpisode, whichEpisode, totalSteps, totalReward)
else:
#print "Evaluation ::\t " + str(totalSteps) + " steps \t" + str(totalReward) + " total reward\t "
print "Evaluation ::\t %d steps \t %.1f total reward" % (totalSteps,
totalReward)
with open('eval_dump.json', 'a') as f:
json.dump(
{
"Steps": totalSteps,
"Episode": whichEpisode,
"Reward": totalReward
}, f)
f.write('\n')
return totalSteps
def run_epoch(epoch, num_steps, prefix):
steps_left = num_steps
while steps_left > 0:
print prefix + " epoch: ", epoch, "steps_left: ", steps_left
terminal = RLGlue.RL_episode(steps_left)
if not terminal:
RLGlue.RL_agent_message("episode_end")
steps_left -= RLGlue.RL_num_steps()
def runEpisode(stepLimit, trial): global whichEpisode terminal=RLGlue.RL_episode(stepLimit) totalSteps=RLGlue.RL_num_steps() totalReward=RLGlue.RL_return() print "Experiment "+str(trial + 1)+"\t Episode "+str(whichEpisode)+"\t "+str(totalSteps)+ " steps \t" + str(totalReward) + " total reward\t " + str(terminal) + " natural end" whichEpisode=whichEpisode+1
def recordTrajectory():
RLGlue.RL_start()
trajectory = []
while True:
roat = RLGlue.RL_step()
trajectory.append(roat)
if roat.terminal:
break
return trajectory
def main():
num_episodes = 10000000
max_steps_per_episode = 50000
RLGlue.RL_init()
for episode in range(0,num_episodes):
RLGlue.RL_episode(max_steps_per_episode)
#print "Episode finished.", time.time()
#print "Score: ", RLGlue.RL_return()
RLGlue.RL_agent_message("save_data data.pkl");
def run_experiment(maxsteps=100, numeps=1):
taskSpec = RLGlue.RL_init()
for ep in range(numeps):
terminal = RLGlue.RL_episode(maxsteps)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
print "Episode " + str(ep) + "\t " + str(
totalSteps) + " steps \t" + str(
totalReward) + " total reward\t " + str(
terminal) + " natural end"
RLGlue.RL_cleanup()
def main():
RLGlue.RL_init()
is_testing = len(sys.argv) > 1 and sys.argv[1] == 'test'
for epoch in xrange(NUM_EPOCHS):
if is_testing:
RLGlue.RL_agent_message("start_testing")
run_epoch(epoch, TEST_LENGTH, "testing")
RLGlue.RL_agent_message("finish_testing " + str(epoch))
else:
run_epoch(epoch, EPOCH_LENGTH, "training")
RLGlue.RL_agent_message("finish_epoch " + str(epoch))
def runEpisode(stepLimit):
# stepLimit of 0 implies no limit
global whichEpisode
terminal = RLGlue.RL_episode(stepLimit)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
print "Episode " + str(whichEpisode) + "\t " + str(
totalSteps) + " steps \t" + str(totalReward) + " total reward\t "
whichEpisode = whichEpisode + 1
def runEpisode(stepLimit):
global whichEpisode
RLGlue.RL_agent_message('reset')
terminal = RLGlue.RL_episode(stepLimit)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
print("Episode " + str(whichEpisode)),
print("\t " + str(totalSteps)),
print(" steps \t" + str(totalReward)),
print " total reward\t " + str(terminal) + " natural end"
RLGlue.RL_agent_message('episode_end')
whichEpisode = whichEpisode + 1
def run(self):
"""
Run the desired number of training epochs, a testing epoch
is conducted after each training epoch.
"""
RLGlue.RL_init()
for epoch in range(1, self.num_epochs + 1):
self.run_epoch(epoch, self.epoch_length, "training")
RLGlue.RL_agent_message("finish_epoch " + str(epoch))
if self.test_length > 0:
RLGlue.RL_agent_message("start_testing")
self.run_epoch(epoch, self.test_length, "testing")
RLGlue.RL_agent_message("finish_testing " + str(epoch))
def runEpisode(is_learning_episode):
global whichEpisode, learningEpisode
RLGlue.RL_episode(0)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
whichEpisode += 1
if is_learning_episode:
learningEpisode += 1
print("Episode " + str(learningEpisode) + "\t " + str(totalSteps) +
" steps \t" + str(totalReward) + " total reward\t ")
else:
print("Evaluation ::\t " + str(totalSteps) + " steps \t" +
str(totalReward) + " total reward\t ")
def demo():
statistics = []
episodeLength = 100
#this_score = evaluateAgent()
#printScore(0, this_score)
#statistics.append(this_score)
for i in range(1, 1000):
RLGlue.RL_env_message("set-start-state " + S)
RLGlue.RL_start()
RLGlue.RL_episode(episodeLength)
this_return = RLGlue.RL_return()
print "%d\t\t%.2f" % (i, this_return)
statistics.append(this_return)
saveResultToCSV(statistics, "MyResults_sarsa1000_ver2.csv")
def run_episode(self): """ Run a single episode """ terminal = RLGlue.RL_episode(10) steps = RLGlue.RL_num_steps() reward = RLGlue.RL_return() #print "Episode %d\t %d steps\t reward: %d" % (episode_number, steps, reward) #print "Episode "+str(episode_number)+"\t "+str(totalSteps)+ " steps \t" + str(totalReward) + " total reward\t " + str(terminal) + " natural end" self.total_reward += reward # Update average x = self.total_reward / (self.episode_number + 1) self.results[self.episode_number] += (x - self.results[self.episode_number]) / self.instance self.episode_number += 1
def agent_step(self, reward, observation):
action = None
self.window.erase()
self.window.addstr('STATE: %s\n' % (observation.intArray))
self.window.addstr('REWARD: %s\n' % (reward))
self.window.addstr('HIT UP, DOWN, LEFT or RIGHT to move...\n')
self.window.refresh()
try:
c = self.window.getch()
if c == curses.KEY_UP:
action = 'N'
elif c == curses.KEY_DOWN:
action = 'S'
elif c == curses.KEY_LEFT:
action = 'W'
elif c == curses.KEY_RIGHT:
action = 'E'
self.window.refresh()
except KeyboardInterrupt:
RLGlue.RL_cleanup()
a = Action()
if action:
a.charArray = [action]
return a
def __init__(self, episodes = 100): """ Initialize experiment """ self.episodes = episodes self.results = [0] * episodes RLGlue.RL_init() self.has_inited = True
def run_epoch(epoch, num_steps, prefix):
""" Run one 'epoch' of training or testing, where an epoch is defined
by the number of steps executed. Prints a progress report after
every trial
Arguments:
num_steps - steps per epoch
prefix - string to print ('training' or 'testing')
"""
steps_left = num_steps
while steps_left > 0:
print prefix + " epoch: ", epoch, "steps_left: ", steps_left
terminal = RLGlue.RL_episode(steps_left)
if not terminal:
RLGlue.RL_agent_message("episode_end")
steps_left -= RLGlue.RL_num_steps()
def start(self):
print "\nExperiment starting!"
taskSpec = RLGlue.RL_init()
print taskSpec
exp_params_for_agent = {}
self.agent_params = self.message_agent(MessageType.exchange_params,
exp_params_for_agent)
# Keep overhead a bit lower by having functions inline
def should_report():
self.step % args.report_freq == 0
def should_evaluate():
step % args.eval_freq == 0 and step > self.agent_params[
'learn_start']
def should_save():
step % args.save_freq == 0
observ_action = RLGlue.RL_start()
while self.step <= self.steps:
observ_action_term = RLGlue.RL_step()
# If game ends, start another
if observ_action_term.terminal:
# Not sure if we need to clean up after every episode, don't think so
RLGlue.RL_start()
self.n_train_episodes += 1
if should_report():
# TODO assert agent steps is equal
print 'Steps: {}'.format(step)
self.message_agent(MessageType.report)
if should_evaluate():
pass
if should_save():
pass
print "A job well done."
RLGlue.RL_cleanup()
def run(self): """ Run the experiment """ if self.has_inited: RLGlue.RL_cleanup() self.instance += 1 self.total_reward = 0 self.episode_number = 0 for i in xrange(self.episodes): self.run_episode()
def evaluateAgent():
sum = 0
sum_of_squares = 0
n = 10
episodeLength = 100
RLGlue.RL_agent_message("freeze learning")
#print "FREEZE LEARNING"
for i in range(0, n):
RLGlue.RL_episode(100)
this_return = RLGlue.RL_return()
sum += this_return
sum_of_squares += this_return**2
mean = sum / n
variance = (sum_of_squares - n * mean * mean) / (n - 1.0)
standard_dev = math.sqrt(variance)
RLGlue.RL_agent_message("unfreeze learning")
#print "UNFREEZE LEARNING"
return mean, standard_dev
def __init__(self, **kwargs):
""" Initialize experiment """
for k, v in kwargs.items():
setattr(self, k, v)
self.returns = [0] * self.episodes
self.steps = [0] * self.episodes
self.po = ProgressOutput()
RLGlue.RL_init()
self.has_inited = True
def offlineDemo():
this_score = evaluateAgent()
printScore(0, this_score)
theFile = open("results.csv", "w")
theFile.close()
if os.path.isfile("Archive.csv"):
os.remove('Archive.csv')
for i in range(0, 200):
for j in range(0, 50):
RLGlue.RL_episode(0)
RLGlue.RL_env_message("stop print")
if j % 20 == 0 and i > 0:
RLGlue.RL_env_message("print")
printScore((i + 1) * 50, this_score)
this_score = evaluateAgent()
printScore((i + 1) * 50, this_score)
theFile = open("results.csv", "a")
theFile.write("%d\t%.2f\t%.2f\n" %
((i) * 50, this_score[0], this_score[1]))
theFile.close()
os.rename('results.csv', 'Archive.csv')
def runEpisode(is_learning_episode):
global whichEpisode, learningEpisode
RLGlue.RL_episode(0)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
whichEpisode += 1
if is_learning_episode:
learningEpisode += 1
print "Episode " + str(learningEpisode) + "\t " + str(totalSteps) + " steps \t" + str(totalReward) + " total reward\t "
else:
print "Evaluation ::\t " + str(totalSteps) + " steps \t" + str(totalReward) + " total reward\t "
# write reward in csv file
list_csv = [str(learningEpisode), str(totalReward)]
f_csv = open('reward.csv', 'a')
writer_r = csv.writer(f_csv, lineterminator = '\n')
writer_r.writerow(list_csv)
f_csv.close()
def runEpisode(is_learning_episode):
global whichEpisode, learningEpisode
RLGlue.RL_episode(0)
totalSteps = RLGlue.RL_num_steps()
totalReward = RLGlue.RL_return()
whichEpisode += 1
if is_learning_episode:
learningEpisode += 1
logger.info("{},{},{},{}".format(
dt.now().strftime("%Y-%m-%d_%H:%M:%S"), learningEpisode,
totalSteps, totalReward))
print "Episode " + str(learningEpisode) + "\t " + str(
totalSteps) + " steps \t" + str(
totalReward) + " total reward\t " + dt.now().strftime(
"%Y%m%d_%H%M%S")
else:
print "Evaluation ::\t " + str(totalSteps) + " steps \t" + str(
totalReward) + " total reward\t "
def main():
whichTrainingMDP = 1
# Uncomment ONE of the following lines to choose your experiment
#loadTetris(whichTrainingMDP); #put the desired parameter set in where MDP is in [0,19]
#loadHelicopter(whichTrainingMDP); #put the desired parameter set in where MDP is in [0,9]
#loadAcrobot(whichTrainingMDP); #put the desired parameter set in where MDP is in [1,49] #0 is standard acrobot
#loadPolyathlon(whichTrainingMDP); #put the desired parameter set in where MDP is in [0,5]
loadMario(True, True, 121, 0, 99, whichTrainingMDP)
# and then,
# just run the experiment:
RLGlue.RL_init()
episodesToRun = 10
totalSteps = 0
for i in range(episodesToRun):
RLGlue.RL_episode(20000)
thisSteps = RLGlue.RL_num_steps()
print "Total steps in episode %d is %d" % (i, thisSteps)
totalSteps += thisSteps
print "Total steps : %d\n" % (totalSteps)
RLGlue.RL_cleanup()
def evaluateAgent():
sum = 0
sum_of_squares = 0
this_return = 0
mean = 0
variance = 0
n = 10
RLGlue.RL_agent_message("freeze learning")
for i in range(0, n):
# We use a cutoff here in case the
#policy is bad and will never end an episode
RLGlue.RL_episode(5000)
this_return = RLGlue.RL_return()
sum += this_return
sum_of_squares += this_return**2
mean = sum / n
variance = (sum_of_squares - n * mean * mean) / (n - 1.0)
standard_dev = math.sqrt(variance)
RLGlue.RL_agent_message("unfreeze learning")
return mean, standard_dev
def offlineDemo():
statistics = []
this_score = evaluateAgent()
printScore(0, this_score)
statistics.append(this_score)
for i in range(0, 20):
for j in range(0, 25):
RLGlue.RL_episode(0)
this_score = evaluateAgent()
printScore((i + 1) * 25, this_score)
statistics.append(this_score)
saveResultToCSV(statistics, "results.csv")
def rlterminate():
RLGlue.doCallWithNoParams(RLGlue.Network.kRLTerm)
