def recommend(
datestrs,
name='departure_tree',
verbose=False,
):
"""
Parameters
---------
datestrs: list of strings
Datestrings of the format YYYY-MM-DD
names: string
The stems fo the filename where the model is store.
verbose: boolean
Returns
-------
recommendations: dict of int
Dictionary keys are datestrings and values are departure times
"""
model_name = name + '.pickle'
try:
# Try to load a saved tree
tree = tools.restore(model_name)
except Exception:
# If unsuccessful, create a new one
tree = create_tree(verbose=verbose)
tools.store(tree, model_name)
features_df = create_features(datestrs)
departures = {}
for datestr in datestrs:
estimated_departure = tree.estimate(features_df.loc[datestr, :])
departures[datestr] = estimated_departure
return departures
def get_trips():
"""
Attempt to restore a saved copy.
If unsuccessful, download a new one.
Returns
-------
trips: list of dictionaries
"""
trips_filename = 'trips.pickle'
try:
trips = tools.restore(trips_filename)
except Exception:
trips = download_data()
tools.store(trips, trips_filename)
return trips
def get_arrival_times(trips_df):
"""
Attempt to restore a saved copy.
if unsuccessful, download a new one.
Parameters
----------
trips_df: DataFrame
Returns
-------
arrival_times_df: DataFrame
"""
arrival_times_filename = 'arrival_times.pickle'
try:
arrival_times_df = tools.restore(arrival_times_filename)
except Exception:
arrival_times_df = None
if arrival_times_df is None:
arrival_times_df = calculate_arrival_times(trips_df)
tools.store(arrival_times_df, arrival_times_filename)
return arrival_times_df
def trainer(env, opt):
agent = tools.qAgent(opt)
sess = tf.InteractiveSession()
obs, brd, Q1 = agent.valueNet()
nobs, nbrd, Q2 = agent.valueNet()
act = tf.placeholder(tf.float32, [None, opt.GAME_SIZE, opt.GAME_SIZE])
rwd = tf.placeholder(tf.float32, [
None,
])
val1 = tf.reduce_sum(tf.reduce_sum(tf.multiply(Q1, act), -1), -1)
val2 = rwd + opt.GAMMA * tf.reduce_max(tf.reduce_max(Q2, -1), 1)
loss = tf.reduce_mean(tf.square(val1 - val2))
trainStep = tf.train.AdamOptimizer(opt.LR).minimize(loss)
sess.run(tf.global_variables_initializer())
saver = tools.restore(sess)
globalStep = 0
for i_e in range(1, opt.MAX_EPISODE + 1):
state = env.reset()
done = False
sumLoss = 0
step = 0
score = 0
while not done:
globalStep += 1
step += 1
if not (globalStep + 1) % opt.EPS_STEP and agent.eps > opt.FIN_EPS:
agent.eps *= opt.EPS_DECAY
agent.obsMem.push(state)
agent.brdMem.push(env.board)
action, actBrd = agent.smpAct(Q1, {
obs: [state],
brd: [env.board]
}, env.board, step)
state, nstate, done, reward = env.step(action)
score += reward
agent.nobsMem.push(nstate)
agent.actMem.push(actBrd)
agent.rwdMem.push(reward)
agent.nbrdMem.push(env.board)
if globalStep >= opt.MEM_SIZE:
randIdx = np.random.choice(opt.MEM_SIZE, opt.BATCH_SIZE)
lossVal, _ = sess.run(
[loss, trainStep],
feed_dict={
obs: agent.obsMem.mem[randIdx],
act: agent.actMem.mem[randIdx],
rwd: agent.rwdMem.mem[randIdx],
nobs: agent.nobsMem.mem[randIdx],
brd: agent.brdMem.mem[randIdx],
nbrd: agent.nbrdMem.mem[randIdx]
})
sumLoss += lossVal
if not i_e % 1:
print(
"====== Episode %d ended with score = %f, avg_loss = %f ======"
% (i_e, score, sumLoss / step))
if i_e > opt.MEM_SIZE and not i_e % 100:
saver.save(sess, 'checkpoints/omok-dqn', global_step=globalStep)
