def get(self, exp_id):
""" Update the parameters for a given experiment
+----------------------------------------------------------------+
| Example |
+================================================================+
|http://example.com/1/setReward.json?key=XXXX&context={'age': 25}|
|&action={'action':'A'}&reward={'click':1} |
+----------------------------------------------------------------+
:param int exp_id: Experiment ID as specified in the url
:param JSON context: The context to train on.
:param JSON action: The action to train on.
:param JSON reward: The reward for the experiment.
:param string key: The key corresponding to the experiment.
:returns: A JSON of the form: {"status":true}
:raises KeyError: 400 Error if Key is not valid.
"""
key = self.get_argument("key", default = False)
__EXP__ = Experiment(exp_id, key)
if __EXP__.is_valid():
context = json.loads(self.get_argument("context", default="{}"))
action = json.loads(self.get_argument("action", default="{}"))
reward = json.loads(self.get_argument("reward", default="{}"))
__EXP__.run_reward_code(context, action, reward)
if self.settings['debug']:
self.write(json.dumps({'status':'success', 'action':action,'context':context, 'reward':reward}))
else:
self.write(json.dumps({'status':'success'}))
else:
self.write_error(400) # Needs proper error handling
def get(self, exp_id):
""" Update the parameters for a given experiment
Input arguments:
exp_id: Experiment ID as specified in the url
context: in JSON get
action: in JSON get
reward: in JSON get
key: in JSON get
Returns:
A JSON object containing "status":true
Or an object containing "error": ...
"""
key = self.get_argument("key", default = False)
__EXP__ = Experiment(exp_id, key)
if __EXP__.is_valid():
context = json.loads(self.get_argument("context", default="{}"))
action = json.loads(self.get_argument("action", default="{}"))
reward = float(self.get_argument("reward", default=0))
__EXP__.run_reward_code(context, action, reward)
if self.settings['debug']:
self.write(json.dumps({'status':'success', 'action':action,'context':context, 'reward':reward}))
else:
self.write(json.dumps({'status':'success'}))
else:
self.write_error(400) # Needs proper error handling
def get(self, exp_id):
""" Update the parameters for a given experiment
+----------------------------------------------------------------+
| Example |
+================================================================+
|http://example.com/1/setReward.json?key=XXXX&context={'age': 25}|
|&action={'action':'A'}&reward={'click':1} |
+----------------------------------------------------------------+
:param int exp_id: Experiment ID as specified in the url
:param JSON context: The context to train on.
:param JSON action: The action to train on.
:param JSON reward: The reward for the experiment.
:param string key: The key corresponding to the experiment.
:returns: A JSON of the form: {"status":true}
:raises KeyError: 400 Error if Key is not valid.
"""
key = self.get_argument("key", default=False)
__EXP__ = Experiment(exp_id, key)
if __EXP__.is_valid():
context = json.loads(self.get_argument("context", default="{}"))
action = json.loads(self.get_argument("action", default="{}"))
reward = json.loads(self.get_argument("reward", default="{}"))
__EXP__.run_reward_code(context, action, reward)
if self.settings['debug']:
self.write(
json.dumps({
'status': 'success',
'action': action,
'context': context,
'reward': reward
}))
else:
self.write(json.dumps({'status': 'success'}))
else:
self.write_error(400) # Needs proper error handling
def advice_time_out():
""" For every experiment, if the advice_id flag is set, we want to
check whether certain advice_id's have timed out according to the
experiment's own settings.
"""
redis_db = Database()
advice_db = Advice()
experiment_ids = redis_db.get_experiment_ids()
for experiment_id in experiment_ids:
# Check experiment properties
exp = Experiment(experiment_id)
if exp.properties["advice_id"] == "True":
# Get all the advices for this experiment
# Check whether or not the date has exceeded the time-out rate
delta_hours = int(exp.properties["delta_hours"])
advices_retrieved = advice_db.advices.find({
"date": {
"$lt": datetime.utcnow() - timedelta(hours=delta_hours)
}
})
for adv in advices_retrieved:
log = exp.get_by_advice_id(str(adv["_id"]))
reward = ast.literal_eval(exp.properties["default_reward"])
exp.run_reward_code(adv["context"], adv["action"], reward)
def get(self, exp_id): # Documentation needs update to advice_id
""" Update the parameters and set a reward for a given experiment.
For parameters, there are two options (next to the mandatory key
and exp_id). The first option is supplying all the information manually,
meaning that you supply the following parameters:
- Context
- Action
- Reward
+----------------------------------------------------------------+
| Example |
+================================================================+
|http://example.com/setreward/EXP_ID?key=KEY&context=CONTEXT |
|&action=ACTION&reward=REWARD |
+----------------------------------------------------------------+
When you have toggled the Advice ID in the experiment properties (second option),
and have received an Advice ID from the getaction call, you only have
to supply the following parameters:
- Advice ID
- Reward
+----------------------------------------------------------------+
| Example |
+================================================================+
|http://example.com/setreward/EXP_ID?key=KEY |
|&advice_id=ADVICE_ID&reward=REWARD |
+----------------------------------------------------------------+
:param int exp_id: Experiment ID as specified in the url.
:param string key: The key corresponding to the experiment.
:param JSON context (optional): The context for the current update.
:param JSON action (optional): The action for the current update.
:param string advice_id (optional): The advice_id for the current \
full update loop.
:param JSON reward: The reward for the current update.
:returns: A JSON indicating success.
:raises 400: If key is not supplied.
:raises 401: If the key or exp_id is invalid.
"""
key = self.get_argument("key", default=False)
if not key:
raise ExceptionHandler(reason="Key not supplied.", status_code=400)
__EXP__ = Experiment(exp_id, key)
if __EXP__.is_valid():
if self.get_argument("advice_id", default="") == "":
context = json.loads(self.get_argument("context",
default="{}"))
action = json.loads(self.get_argument("action", default="{}"))
else:
advice_id = self.get_argument("advice_id", default="")
log = __EXP__.get_by_advice_id(advice_id)
if log == False:
self.finish("Advice ID does not exist!")
else:
context = log['context']
action = log['action']
reward = json.loads(self.get_argument("reward", default="{}"))
__EXP__.run_reward_code(context, action, reward)
__EXP__.log_setreward_data(context, action, reward)
if self.settings['debug']:
self.write(
json.dumps({
'status': 'success',
'action': action,
'context': context,
'reward': reward
}))
else:
self.write(json.dumps({'status': 'success'}))
else:
raise ExceptionHandler(reason="Key or exp_id is invalid.",
status_code=401)
def get(self, exp_id):
""" Simulate your experiment based on four scripts, which create a closed feedback loop.
+--------------------------------------------------------------------+
| Example |
+====================================================================+
|http://example.com/eval/EXP_ID/simulate?N=1000&log_stats=True |
|&verbose=True&seed=10 |
+--------------------------------------------------------------------+
:requires: A secure cookie, obtained by logging in.
:param int exp_id: Experiment ID as specified in the url.
:param int N: The number of simulation draws.
:param bool log_stats: Flag for logging the results in the database (default is False)
:param bool verbose: Flag for displaying the results in the returning JSON object (default is True)
:param int seed (optional): Set numpy seed.
:returns: A JSON indicating success when verbose flag is False, and a JSON with all the data when verbose flag is True.
:raises 400: If the experiment does not belong to this user or the exp_id is wrong.
:raises 401: If user is not logged in or if there is no secure cookie available.
"""
if self.get_current_user():
if self.validate_user_experiment(exp_id):
N = int(self.get_argument("N", default=1000))
log_stats = self.get_argument("log_stats", default=False)
verbose = self.get_argument("verbose", default=True)
seed = self.get_argument("seed", default=None)
if seed is None:
seed = np.random.randint(2**32 - 1, dtype=np.uint32)
if verbose == "True":
verbose = True
else:
verbose = False
if log_stats == "True":
log_stats = True
else:
log_stats = False
__EXP__ = Experiment(exp_id)
data = {}
with self.temp_seed(int(seed)):
for i in range(N):
# Generate context
context = __EXP__.run_context_code()
# Get action
action = __EXP__.run_action_code(context)
# Generate reward
reward = __EXP__.run_get_reward_code(context, action)
# Set reward
__EXP__.run_reward_code(context, action, reward)
# Get theta
theta = __EXP__.get_theta()
# Save stats
data[str(i)] = {
'context': context.copy(),
'action': action.copy(),
'reward': reward.copy(),
'theta': theta.copy()
}
context.clear()
action.clear()
reward.clear()
if log_stats == True:
__EXP__.log_simulation_data(data.copy())
data_tmp = data.copy()
data.clear()
if verbose == True:
self.write(
json.dumps({
'simulate': 'success',
'experiment': exp_id,
'data': data_tmp
}))
else:
self.write(
json.dumps({
'simulate': 'success',
'experiment': exp_id,
'theta': theta
}))
else:
raise ExceptionHandler(
reason="Experiment could not be validated.",
status_code=401)
else:
raise ExceptionHandler(reason="Could not validate user.",
status_code=401)
def get(self, exp_id):
""" Simulate your experiment on a simple model
The model that is drawn from is:
y = -(x - c)**2 + c2 + rnorm(mu,var)
Currently there is no context. Make sure that the action of your
experiment results in:
{"x" : x}
This is how the model currently expects your action to be formulated.
This might become more flexible later on.
+--------------------------------------------------------------------+
| Example |
+====================================================================+
|http://example.com/eval/5/simulate?key=XXX&N=10&c=5&c2=10&mu=0&var=1|
+--------------------------------------------------------------------+
:param int exp_id: Experiment ID as specified in the url
:param string key: The key corresponding to the experiment
:param int N: The number of simulation draws
:param int c: The size of the parabola
:param int c2: The height of the parabola
:param int mu: The mean of the noise on the model
:param int var: The variance of the noise on the model
:param string log_stats: Flag for logging the results in the database
:returns: A JSON of the form: {"simulate":"success"}
:raises AuthError: 401 Invalid Key
"""
key = self.get_argument("key", default = False)
# Number of draws
N = int(self.get_argument("N", default = 1000))
log_stats = self.get_argument("log_stats", default = True)
# Parameterset for the simulator
c = float(self.get_argument("c", default = 5))
c2 = float(self.get_argument("c2", default = 10))
mu = float(self.get_argument("mu", default = 0))
var = float(self.get_argument("var", default = .1))
if not key:
self.set_status(401)
self.write("Key not given")
return
__EXP__ = Experiment(exp_id, key)
rewards = np.array([0])
reward_over_time = np.array([])
regret = np.array([0])
if __EXP__.is_valid():
for i in range(N):
# Generate context
context = {}
# Get action
action = __EXP__.run_action_code(context)
# Generate reward
y = -(action["x"] - c)**2 + c2 + np.random.normal(mu, var)
#y = 15 + 8*action["x"] + 10*action["x"]**2 + np.random.normal(mu, var)
reward = {"y" : y}
# Set reward
__EXP__.run_reward_code(context, action, reward)
# Save stats
rewards = np.append(rewards, y)
tmp_rot = (rewards[-1] + y) / (i+1)
reward_over_time = np.append(reward_over_time, tmp_rot)
regret = np.append(regret, (regret[-1] + (c2 - y)))
#self.write("n = {}, Regret is: {}, reward = {} <br>".format(i,regret[-1], rewards[-1]))
# Now save the data together with a timestamp in the logs
# To read out the Numpy array data out again, use array =
# pickle.loads(record['feature'])
# FOR FUTURE, the json_tricks package might be interesting
if log_stats == True:
print("Logging data")
__EXP__.log_data({
"type" : "evaluation",
"time" : int(time.time()),
"experiment" : exp_id,
"N" : N,
"c" : c,
"c2" : c2,
"rewards" : Binary(_pickle.dumps(rewards, protocol = 2), subtype = 128),
"reward_over_time" : Binary(_pickle.dumps(reward_over_time, protocol = 2), subtype = 128),
"regret" : Binary(_pickle.dumps(regret, protocol = 2), subtype = 128)
})
self.write(json.dumps({'simulate':'success','experiment':exp_id}))
else:
self.set_status(401)
self.write("Key is not valid for this experiment")
return
