def main(debug):
interval = choice(INTERVALS)
minutes = 0
while True:
minutes += 1
seconds_to_run = 60 * minutes
seconds_info_intervals = seconds_to_run / 4
logging.error('Training each currency for {0} minutes'.format(minutes))
# shuffle(CURRENCIES)
for currency, min_trail in CURRENCIES.iteritems():
pip_mul = 100. if 'JPY' in currency else 10000.
actions = calculateActions(min_trail)
df = loadData(currency, interval)
df = df[-2000:]
df = getBackgroundKnowledge(df, PERIODS)
# print df
# break
alpha = 0
epsilon = 0
q = loadQ(currency, interval)
time_start = time()
time_interval = 0
epoch = 0
rewards = []
errors = []
ticks = []
logging.warn('Training {0} on {1} with {2} ticks...'.format(currency, interval, len(df)))
while True:
epoch += 1
logging.info(' ')
logging.info('{0}'.format('=' * 20))
logging.info('EPOCH {0}'.format(epoch))
index_start = randint(0, len(df)-20)
df_inner = df.iloc[index_start:]
logging.info('Epoch: at {0} with {1} ticks'.format(index_start, len(df_inner)))
q, r, error, tick = train(df_inner, q, alpha, epsilon, PERIODS, actions, pip_mul)
# results
rewards.append(r)
errors.append(error * pip_mul)
ticks.append(tick)
# win ratio
wins = [1. if r > 0. else 0. for r in rewards]
win_ratio = np.mean(wins)
# logging.error('wr {0}'.format(win_ratio))
# adjust values
epsilon = np.sqrt((1 - win_ratio) * 100.) / 100.
alpha = epsilon / 2.
# logging.error('new alpha = {0}'.format(alpha))
if time() - time_start > time_interval or debug:
# prune lengths
while len(rewards) > 1000 + minutes * 1000:
rewards.pop(0)
errors.pop(0)
ticks.pop(0)
# RMSD
rmsd = np.sqrt(np.mean([e**2 for e in errors]))
logging.info('RMSD: {0}'.format(rmsd))
logging.warn('{0} [{1:05d}] RMSD {2:03d} PPT {3:03d} WR {5:.0f}% [ticks:{6:.1f} sum:{4:.1f}, a:{7:.2f}, e:{8:.2f}]'.format(
currency,
epoch,
int(rmsd),
int(np.mean(rewards) * pip_mul),
sum(rewards),
np.mean(wins) * 100,
np.mean(ticks),
alpha * 100,
epsilon * 100,
))
# exit
if (time() - time_start >= seconds_to_run) or debug:
break
# saveQ(currency, interval, q)
time_interval += seconds_info_intervals
saveQ(currency, interval, q)
summarizeActions(q)
if debug:
break # currencies
if debug:
break # forever
def main(debug):
interval = choice(INTERVALS)
minutes = 0
while True:
minutes += 1
seconds_to_run = 60 * minutes
seconds_info_intervals = seconds_to_run / 4
logging.error('Training each currency for {0} minutes'.format(minutes))
# shuffle(CURRENCIES)
for currency, min_trail in CURRENCIES.iteritems():
pip_mul = 100. if 'JPY' in currency else 10000.
actions = calculateActions(min_trail)
df = loadData(currency, interval)
df = df[-2000:]
df = getBackgroundKnowledge(df, PERIODS)
# print df
# break
alpha = 0
epsilon = 0
q = loadQ(currency, interval)
time_start = time()
time_interval = 0
epoch = 0
rewards = []
errors = []
ticks = []
logging.warn('Training {0} on {1} with {2} ticks...'.format(
currency, interval, len(df)))
while True:
epoch += 1
logging.info(' ')
logging.info('{0}'.format('=' * 20))
logging.info('EPOCH {0}'.format(epoch))
index_start = randint(0, len(df) - 20)
df_inner = df.iloc[index_start:]
logging.info('Epoch: at {0} with {1} ticks'.format(
index_start, len(df_inner)))
q, r, error, tick = train(df_inner, q, alpha, epsilon, PERIODS,
actions, pip_mul)
# results
rewards.append(r)
errors.append(error * pip_mul)
ticks.append(tick)
# win ratio
wins = [1. if r > 0. else 0. for r in rewards]
win_ratio = np.mean(wins)
# logging.error('wr {0}'.format(win_ratio))
# adjust values
epsilon = np.sqrt((1 - win_ratio) * 100.) / 100.
alpha = epsilon / 2.
# logging.error('new alpha = {0}'.format(alpha))
if time() - time_start > time_interval or debug:
# prune lengths
while len(rewards) > 1000 + minutes * 1000:
rewards.pop(0)
errors.pop(0)
ticks.pop(0)
# RMSD
rmsd = np.sqrt(np.mean([e**2 for e in errors]))
logging.info('RMSD: {0}'.format(rmsd))
logging.warn(
'{0} [{1:05d}] RMSD {2:03d} PPT {3:03d} WR {5:.0f}% [ticks:{6:.1f} sum:{4:.1f}, a:{7:.2f}, e:{8:.2f}]'
.format(
currency,
epoch,
int(rmsd),
int(np.mean(rewards) * pip_mul),
sum(rewards),
np.mean(wins) * 100,
np.mean(ticks),
alpha * 100,
epsilon * 100,
))
# exit
if (time() - time_start >= seconds_to_run) or debug:
break
# saveQ(currency, interval, q)
time_interval += seconds_info_intervals
saveQ(currency, interval, q)
summarizeActions(q)
if debug:
break # currencies
if debug:
break # forever
def main(debug):
minutes = 0
while True:
minutes += 1
seconds_to_run = 60 * minutes
seconds_info_intervals = seconds_to_run / 5
logging.error('Training each currency for {0} minutes'.format(minutes))
shuffle(DATA)
for info in DATA:
logging.debug('Currency info: {0}'.format(info))
currency = info['currency']
interval = info['intervals'][0]
pip_mul = info['pip_mul']
df = loadData(currency, interval, 'train')
df = getBackgroundKnowledge(df, PERIODS)
alpha = 0.
epsilon = alpha / 2.
q = loadQ(currency, interval)
time_start = time()
time_interval = seconds_info_intervals
epoch = 0
rewards = []
errors = []
ticks = []
logging.warn('Training {0} on {1} with {2} ticks [m:{3}]'.format(
currency,
interval,
len(df),
minutes,
))
while True:
epoch += 1
logging.info(' ')
logging.info('{0}'.format('=' * 20))
logging.info('EPOCH {0}'.format(epoch))
index_start = randint(0, len(df)-20)
df_inner = df.iloc[index_start:]
logging.info('Epoch: at {0} with {1} ticks'.format(index_start, len(df_inner)))
q, r, error, tick = train(df_inner, q, alpha, epsilon, PERIODS, ACTIONS, pip_mul, info['std'])
# results
error *= pip_mul
rewards.append(r)
errors.append(error)
ticks.append(tick)
# win ratio
wins = [1. if r > 0. else 0. for r in rewards]
win_ratio = np.mean(wins)
# logging.error('wr {0}'.format(win_ratio))
# adjust values
alpha = 1.0102052281586786e+000 + (-2.0307383627607809e+000 * win_ratio) + (1.0215546892913909e+000 * win_ratio**2)
epsilon = 3.9851080604500078e-001 + (2.1874724815820201e-002 * win_ratio) + (-4.1444101741886652e-001 * win_ratio**2)
# logging.error('new alpha = {0}'.format(alpha))
# only do updates at interval
if time() - time_start > time_interval or debug:
# prune lengths
while len(rewards) > 1000 + minutes * 1000:
rewards.pop(0)
errors.pop(0)
ticks.pop(0)
# RMSD
rmsd = np.sqrt(np.mean([e**2 for e in errors]))
# logging.error('RMSD: {0} from new error {1}'.format(rmsd, error))
logging.warn('{0} [{1:05d}] RMSD {2:03d} PPT {3:03d} WR {5:.0f}% [ticks:{6:.1f} sum:{4:.1f}, a:{7:.2f}, e:{8:.2f}]'.format(
currency,
epoch,
int(rmsd),
int(np.mean(rewards) * pip_mul),
sum(rewards),
np.mean(wins) * 100,
np.mean(ticks),
alpha * 100,
epsilon * 100,
))
# exit
if time_interval >= seconds_to_run or debug:
break
# continue
time_interval += seconds_info_intervals
saveQ(currency, interval, q)
saveQ(currency, interval, q)
summarizeActions(q)
if debug:
break # currencies
if debug:
break # forever
def main(debug):
minutes = 0
while True:
minutes += 1
seconds_to_run = 60 * minutes
seconds_info_intervals = seconds_to_run / 5
logging.error('Training each currency for {0} minutes'.format(minutes))
shuffle(DATA)
for info in DATA:
logging.debug('Currency info: {0}'.format(info))
currency = info['currency']
interval = info['intervals'][0]
pip_mul = info['pip_mul']
df = loadData(currency, interval, 'train')
df = getBackgroundKnowledge(df, PERIODS)
alpha = 0.
epsilon = alpha / 2.
q = loadQ(currency, interval)
time_start = time()
time_interval = seconds_info_intervals
epoch = 0
rewards = []
errors = []
ticks = []
logging.warn('Training {0} on {1} with {2} ticks [m:{3}]'.format(
currency,
interval,
len(df),
minutes,
))
while True:
epoch += 1
logging.info(' ')
logging.info('{0}'.format('=' * 20))
logging.info('EPOCH {0}'.format(epoch))
index_start = randint(0, len(df) - 20)
df_inner = df.iloc[index_start:]
logging.info('Epoch: at {0} with {1} ticks'.format(
index_start, len(df_inner)))
q, r, error, tick = train(df_inner, q, alpha, epsilon, PERIODS,
ACTIONS, pip_mul, info['std'])
# results
error *= pip_mul
rewards.append(r)
errors.append(error)
ticks.append(tick)
# win ratio
wins = [1. if r > 0. else 0. for r in rewards]
win_ratio = np.mean(wins)
# logging.error('wr {0}'.format(win_ratio))
# adjust values
alpha = 1.0102052281586786e+000 + (
-2.0307383627607809e+000 *
win_ratio) + (1.0215546892913909e+000 * win_ratio**2)
epsilon = 3.9851080604500078e-001 + (
2.1874724815820201e-002 *
win_ratio) + (-4.1444101741886652e-001 * win_ratio**2)
# logging.error('new alpha = {0}'.format(alpha))
# only do updates at interval
if time() - time_start > time_interval or debug:
# prune lengths
while len(rewards) > 1000 + minutes * 1000:
rewards.pop(0)
errors.pop(0)
ticks.pop(0)
# RMSD
rmsd = np.sqrt(np.mean([e**2 for e in errors]))
# logging.error('RMSD: {0} from new error {1}'.format(rmsd, error))
logging.warn(
'{0} [{1:05d}] RMSD {2:03d} PPT {3:03d} WR {5:.0f}% [ticks:{6:.1f} sum:{4:.1f}, a:{7:.2f}, e:{8:.2f}]'
.format(
currency,
epoch,
int(rmsd),
int(np.mean(rewards) * pip_mul),
sum(rewards),
np.mean(wins) * 100,
np.mean(ticks),
alpha * 100,
epsilon * 100,
))
# exit
if time_interval >= seconds_to_run or debug:
break
# continue
time_interval += seconds_info_intervals
saveQ(currency, interval, q)
saveQ(currency, interval, q)
summarizeActions(q)
if debug:
break # currencies
if debug:
break # forever