def processAnswer(self, exp_uid, args_json, db, ell):
"""
reporting back the reward of pulling the arm suggested by getQuery
Expected input (in json structure with string keys):
(str) query_uid : unique identifier of query
(int) index_winner : index of arm must be {index_left,index_right}
Expected output (comma separated):
if error:
return (JSON) '{}', (bool) False, (str) error
else:
return (JSON) '{}', (bool) True,''
Usage:
processAnswer_args_json,didSucceed,message = app.processAnswer(db_API,exp_uid,processAnswer_args_json)
Example input:
processAnswer_args_json = {"query_uid": "4d02a9924f92138287edd17ca5feb6e1", "index_winner": 8}
Example output:
processAnswer_response_json = {}
"""
try:
app_id = self.app_id
log_entry = {
'exp_uid': exp_uid,
'task': 'processAnswer',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.processAnswer input args_json is in improper format" % self.app_id
return '{}', False, error
# check for the fields that must be contained in args or error occurs
necessary_fields = ['index_winner', 'query_uid']
for field in necessary_fields:
try:
args_dict[field]
except KeyError:
error = "%s.processAnswer input arguments missing field: %s" % (
self.app_id, str(field))
return '{}', False, error
# get list of algorithms associated with project
alg_list, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'alg_list')
# get alg_id
query_uid = args_dict['query_uid']
alg_uid, didSucceed, message = db.get(app_id + ':queries',
query_uid, 'alg_uid')
for algorithm in alg_list:
if alg_uid == algorithm['alg_uid']:
alg_id = algorithm['alg_id']
alg_label = algorithm['alg_label']
num_reported_answers, didSucceed, message = db.increment(
app_id + ':experiments', exp_uid,
'num_reported_answers_for_' + alg_uid)
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
targets, didSucceed, message = db.get(app_id + ':queries',
query_uid, 'target_indices')
for target in targets:
if target['label'] == 'left':
index_left = target['index']
if target['label'] == 'right':
index_right = target['index']
if target['flag'] == 1:
index_painted = target['index']
index_winner = args_dict['index_winner']
# update query doc
timestamp_query_generated, didSucceed, message = db.get(
app_id + ':queries', query_uid, 'timestamp_query_generated')
datetime_query_generated = utils.str2datetime(
timestamp_query_generated)
timestamp_answer_received = args_dict.get('meta', {}).get(
'timestamp_answer_received', None)
if timestamp_answer_received == None:
datetime_answer_received = datetime_query_generated
else:
datetime_answer_received = utils.str2datetime(
timestamp_answer_received)
delta_datetime = datetime_answer_received - datetime_query_generated
round_trip_time = delta_datetime.seconds + delta_datetime.microseconds / 1000000.
response_time = float(args_dict.get('response_time', 0.))
db.set(app_id + ':queries', query_uid, 'response_time',
response_time)
db.set(app_id + ':queries', query_uid, 'network_delay',
round_trip_time - response_time)
db.set(app_id + ':queries', query_uid, 'index_winner',
index_winner)
# call processAnswer
didSucceed, dt = utils.timeit(alg.processAnswer)(
resource=rc,
index_left=index_left,
index_right=index_right,
index_painted=index_painted,
index_winner=index_winner)
log_entry_durations = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'processAnswer',
'duration': dt
}
log_entry_durations.update(rc.getDurations())
meta = {'log_entry_durations': log_entry_durations}
###############
predict_id = 'arm_ranking'
params = {'alg_label': alg_label}
predict_args_dict = {'predict_id': predict_id, 'params': params}
predict_args_json = json.dumps(predict_args_dict)
db.submit_job(app_id,
exp_uid,
'predict',
predict_args_json,
ignore_result=True)
###############
response_args_dict = {}
args_out = {'args': response_args_dict, 'meta': meta}
response_json = json.dumps(args_out)
log_entry = {
'exp_uid': exp_uid,
'task': 'processAnswer',
'json': response_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return response_json, True, ""
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'processAnswer',
'error': error,
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error
def predict(self, exp_uid, args_json, db, ell):
"""
uses current model empirical estimates to forecast which index has the highest mean
Expected input (in json structure with string keys):
(string) predict_id : identifier for the desired prediction
(list) params : dictionary of stat_id specific fields.
##########
Description: uses current model empirical estimates to forecast the ranking of the arms in order of likelhood of being the best from most to least likely
Expected input:
(string) predict_id : 'arm_ranking'
(dict) params : dictionary with fields
(string) alg_label : describes target algorithm to use
Expected output (in json structure):
(list of dicts with fields):
(int) index : index of target
(int) rank : rank that the algorithm assigns to index (rank 0 is most likely the best arm)
"""
try:
app_id = self.app_id
log_entry = {
'exp_uid': exp_uid,
'task': 'predict',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.predict failed to convert input args_json due to improper format" % (
self.app_id)
return '{}', False, error
# check for the fields that must be contained in args or error occurs
necessary_fields = ['predict_id', 'params']
for field in necessary_fields:
try:
args_dict[field]
except KeyError:
error = "%s.predict input arguments missing field: %s" % (
self.app_id, str(field))
return '{}', False, error
predict_id = args_dict['predict_id']
params = args_dict['params']
if predict_id == "arm_ranking":
alg_label = params['alg_label']
# get list of algorithms associated with project
alg_list, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'alg_list')
# get alg_id
for algorithm in alg_list:
if alg_label == algorithm['alg_label']:
alg_id = algorithm['alg_id']
alg_uid = algorithm['alg_uid']
num_reported_answers, didSucceed, message = db.get(
app_id + ':experiments', exp_uid,
'num_reported_answers_for_' + alg_uid)
if type(num_reported_answers) != int:
num_reported_answers = 0
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
# call getQuery
scores, precisions, dt = utils.timeit(alg.predict)(resource=rc)
log_entry_durations = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'predict',
'duration': dt
}
log_entry_durations.update(rc.getDurations())
meta = {'log_entry_durations': log_entry_durations}
import numpy
ranks = (-numpy.array(scores)).argsort().tolist()
n = len(scores)
indexes = numpy.array(range(n))[ranks]
scores = numpy.array(scores)[ranks]
precisions = numpy.array(precisions)[ranks]
ranks = range(n)
targets = []
for index in range(n):
targets.append({
'index': indexes[index],
'rank': ranks[index],
'score': scores[index],
'precision': precisions[index]
})
log_entry = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'timestamp': utils.datetimeNow()
}
log_entry.update({
'targets': targets,
'num_reported_answers': num_reported_answers
})
ell.log(app_id + ':ALG-EVALUATION', log_entry)
response_args_dict = {}
args_out = {'args': response_args_dict, 'meta': meta}
predict_json = json.dumps(args_out)
log_entry = {
'exp_uid': exp_uid,
'task': 'predict',
'json': predict_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return predict_json, True, ''
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'predict',
'error': str(error),
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
didSucceed, message = ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error
def getQuery(self, exp_uid, args_json, db, ell):
"""
A request to ask which two arms to duel next
Expected input (in jsonstructure with string keys):
[optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid.
Expected output (in json structure with string keys):
(list) target_indices : list that stores dictionary of targets with fields:
{
(int) index : the index of the target of relevance
(str) label : in {'left','right'} for display
(int) flag : integer for algorithm's use
}
(str) query_uid : unique identifier of query (used to look up for processAnswer)
Usage:
getQuery_response_json,didSucceed,message = app.getQuery(db_API,exp_uid,getQuery_args_json)
Example input:
getQuery_args_json = {"participant_uid": "0077110d03cf06b8f77d11acc399e8a7"}
Example output:
getQuery_response_json = {"query_uid": "4d02a9924f92138287edd17ca5feb6e1", "target_indices": [{"index": 9, "flag": 0, "label": "left"}, {"index": 8, "flag": 1, "label": "right"}]}
"""
try:
app_id = self.app_id
log_entry = {
'exp_uid': exp_uid,
'task': 'getQuery',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.initExp input args_json is in improper format" % self.app_id
return '{}', False, error
# get list of algorithms associated with project
alg_list, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'alg_list')
alg_label_to_alg_id = {}
alg_label_to_alg_uid = {}
for algorithm in alg_list:
alg_label_to_alg_id[
algorithm['alg_label']] = algorithm['alg_id']
alg_label_to_alg_uid[
algorithm['alg_label']] = algorithm['alg_uid']
algorithm_management_settings, didSucceed, message = db.get(
app_id + ':experiments', exp_uid,
'algorithm_management_settings')
# ASSIGN ALGORITHM TO PARTICIPANT
if 'participant_uid' in args_dict:
participant_uid = args_dict['participant_uid']
else:
participant_uid = exp_uid
participant_doc_exists, didSucceed, message = db.exists(
app_id + ':participants', participant_uid, 'participant_uid')
first_participant_query = not participant_doc_exists
if first_participant_query:
db.set(app_id + ':participants', participant_uid,
'participant_uid', participant_uid)
db.set(app_id + ':participants', participant_uid, 'exp_uid',
exp_uid)
participant_to_algorithm_management, didSucceed, message = db.get(
app_id + ':experiments', exp_uid,
'participant_to_algorithm_management')
if (participant_uid == exp_uid) or (
participant_to_algorithm_management
== 'one_to_many') or (first_participant_query):
if algorithm_management_settings[
'mode'] == 'fixed_proportions':
proportions_list = algorithm_management_settings['params'][
'proportions']
prop = [
prop_item['proportion']
for prop_item in proportions_list
]
prop_item = numpy.random.choice(alg_list, p=prop)
else:
raise Exception('algorithm_management_mode : ' +
algorithm_management_settings['mode'] +
' not implemented')
alg_id = alg_label_to_alg_id[prop_item['alg_label']]
alg_uid = alg_label_to_alg_uid[prop_item['alg_label']]
alg_label = prop_item['alg_label']
if (first_participant_query) and (
participant_to_algorithm_management == 'one_to_one'):
db.set(app_id + ':participants', participant_uid, 'alg_id',
alg_id)
db.set(app_id + ':participants', participant_uid,
'alg_uid', alg_uid)
elif (participant_to_algorithm_management == 'one_to_one'):
# If here, then alg_uid should already be assigned in participant doc
alg_id, didSucceed, message = db.get(app_id + ':participants',
participant_uid, 'alg_id')
alg_uid, didSucceed, message = db.get(app_id + ':participants',
participant_uid,
'alg_uid')
else:
raise Exception('participant_to_algorithm_management : ' +
participant_to_algorithm_management +
' not implemented')
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
# call getQuery
index_left, index_right, index_painted, dt = utils.timeit(
alg.getQuery)(resource=rc)
# check for context
context_type, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'context_type')
context, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'context')
# log
log_entry_durations = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'getQuery',
'duration': dt
}
log_entry_durations.update(rc.getDurations())
meta = {'log_entry_durations': log_entry_durations}
# create JSON query payload
if index_left == index_painted:
targets = [{
'index': index_left,
'label': 'left',
'flag': 1
}, {
'index': index_right,
'label': 'right',
'flag': 0
}]
else:
targets = [{
'index': index_left,
'label': 'left',
'flag': 0
}, {
'index': index_right,
'label': 'right',
'flag': 1
}]
timestamp = str(utils.datetimeNow())
query_uid = utils.getNewUID()
query = {}
query['query_uid'] = query_uid
query['target_indices'] = targets
# save query data to database
query_doc = {}
query_doc.update(query)
query_doc['participant_uid'] = participant_uid
query_doc['alg_uid'] = alg_uid
query_doc['exp_uid'] = exp_uid
query_doc['alg_label'] = alg_label
query_doc['timestamp_query_generated'] = timestamp
db.set_doc(app_id + ':queries', query_uid, query_doc)
# add context after updating query doc to avoid redundant information
query['context_type'] = context_type
query['context'] = context
args_out = {'args': query, 'meta': meta}
response_json = json.dumps(args_out)
log_entry = {
'exp_uid': exp_uid,
'task': 'getQuery',
'json': response_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return response_json, True, ''
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'getQuery',
'error': error,
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error
def daemonProcess(self, exp_uid, args_json, db, ell):
try:
app_id = self.app_id
log_entry = {
'exp_uid': exp_uid,
'task': 'daemonProcess',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.daemonProcess input args_json is in improper format" % self.app_id
return '{}', False, error
# check for the fields that must be contained in args or error occurs
necessary_fields = ['alg_uid', 'daemon_args']
for field in necessary_fields:
try:
args_dict[field]
except KeyError:
error = "%s.daemonProcess input arguments missing field: %s" % (
self.app_id, str(field))
return '{}', False, error
alg_daemon_args = args_dict['daemon_args']
alg_uid = args_dict['alg_uid']
alg_id, didSucceed, message = db.get(app_id + ':algorithms',
alg_uid, 'alg_id')
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
# rc = ResourceClient(app_id,exp_uid,alg_uid,db)
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
didSucceed, dt = utils.timeit(alg.daemonProcess)(
resource=rc, daemon_args_dict=alg_daemon_args)
log_entry_durations = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'daemonProcess',
'duration': dt
}
log_entry_durations.update(rc.getDurations())
meta = {'log_entry_durations': log_entry_durations}
daemon_message = {}
args_out = {'args': daemon_message, 'meta': meta}
response_json = json.dumps(args_out)
log_entry = {
'exp_uid': exp_uid,
'task': 'daemonProcess',
'json': response_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return response_json, True, ''
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'daemonProcess',
'error': error,
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error
def initExp(self, exp_uid, args_json, db, ell):
"""
initialize the project and necessary experiments
Expected input (in json structure with string keys):
(int) n: number of arms
(float) failure_probability : confidence
[optional] (list of dicts) alg_list : with fields (Defaults given by Info.get_app_default_alg_list)
(string) alg_id : valid alg_id for this app_id
(string) alg_label : unique identifier for algorithm (e.g. may have experiment with repeated alg_id's, but alg_labels must be unqiue, will also be used for plot legends
[optional] (string) test_alg_label : must be one of the alg_label's in alg_list (Default is self)
[optional] (dict) algorithm_management_settings : dictionary with fields (string) 'mode' and (dict) 'params'. mode in {'pure_exploration','explore_exploit','fixed_proportions'}. Default is 'fixed_proportions' and allocates uniform probability to each algorithm. If mode=fixed_proportions then params is a dictionary that contains the field 'proportions' which is a list of dictionaries with fields 'alg_label' and 'proportion' for all algorithms in alg_list. All proportions must be positive and sum to 1 over all algs in alg_list
[optional] (string) participant_to_algorithm_management : in {'one_to_one','one_to_many'}. Default is 'one_to_many'.
[optional] (string) instructions
[optional] (string) debrief
[optional] (int) num_tries
Expected output:
if error:
return (JSON) '{}', (bool) False, (str) error_str
else:
return (JSON) '{}', (bool) True,''
Usage:
initExp_response_json,didSucceed,message = app.initExp(db_API,exp_uid,initExp_args_json)
Example input:
initExp_args_json = {"participant_to_algorithm_management": "one_to_many", "alg_list": [{"alg_label": "BR_LilUCB", "alg_id": "BR_LilUCB", "params": {}}], "algorithm_management_settings": {"params": {"proportions": [{"alg_label": "BR_LilUCB", "proportion": 1.0}]}, "mode": "fixed_proportions"}, "failure_probability": 0.01, "n": 10}
Example output:
initExp_response_json = {}
"""
try:
app_id = self.app_id
# remove any reminants of an experiment if it exists
didSucceed, message = db.delete_docs_with_filter(
'experiments_admin', {'exp_uid': exp_uid})
didSucceed, message = db.delete_docs_with_filter(
app_id + ':experiments', {'exp_uid': exp_uid})
didSucceed, message = db.delete_docs_with_filter(
app_id + ':queries', {'exp_uid': exp_uid})
didSucceed, message = db.delete_docs_with_filter(
app_id + ':participants', {'exp_uid': exp_uid})
didSucceed, message = db.delete_docs_with_filter(
app_id + ':algorithms', {'exp_uid': exp_uid})
didSucceed, message = ell.delete_logs_with_filter(
app_id + ':APP-CALL', {'exp_uid': exp_uid})
didSucceed, message = ell.delete_logs_with_filter(
app_id + ':APP-RESPONSE', {'exp_uid': exp_uid})
didSucceed, message = ell.delete_logs_with_filter(
app_id + ':APP-EXCEPTION', {'exp_uid': exp_uid})
didSucceed, message = ell.delete_logs_with_filter(
app_id + ':ALG-DURATION', {'exp_uid': exp_uid})
didSucceed, message = ell.delete_logs_with_filter(
app_id + ':ALG-EVALUATION', {'exp_uid': exp_uid})
# add indexes (only adds them if they do not already exist)
didSucceed, message = db.ensure_index('experiments_admin',
{'exp_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':experiments',
{'exp_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':queries',
{'query_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':queries',
{'exp_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':queries',
{'alg_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':queries',
{'participant_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':participants',
{'participant_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':participants',
{'exp_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':algorithms',
{'alg_uid': 1})
didSucceed, message = db.ensure_index(app_id + ':algorithms',
{'exp_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-CALL',
{'exp_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-CALL',
{'timestamp': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-CALL', {
'exp_uid': 1,
'timestamp': 1
})
didSucceed, message = ell.ensure_index(app_id + ':APP-CALL', {
'exp_uid': 1,
'task': 1
})
didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE',
{'exp_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE',
{'timestamp': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE', {
'exp_uid': 1,
'timestamp': 1
})
didSucceed, message = ell.ensure_index(app_id + ':APP-RESPONSE', {
'exp_uid': 1,
'task': 1
})
didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION',
{'exp_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION',
{'timestamp': 1})
didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION', {
'exp_uid': 1,
'timestamp': 1
})
didSucceed, message = ell.ensure_index(app_id + ':APP-EXCEPTION', {
'exp_uid': 1,
'task': 1
})
didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION',
{'exp_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION',
{'alg_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION',
{'timestamp': 1})
didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', {
'exp_uid': 1,
'timestamp': 1
})
didSucceed, message = ell.ensure_index(app_id + ':ALG-DURATION', {
'alg_uid': 1,
'task': 1
})
didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION',
{'exp_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION',
{'alg_uid': 1})
didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION',
{'timestamp': 1})
didSucceed, message = ell.ensure_index(app_id + ':ALG-EVALUATION',
{
'exp_uid': 1,
'timestamp': 1
})
db.set('experiments_admin', exp_uid, 'exp_uid', exp_uid)
db.set('experiments_admin', exp_uid, 'app_id', app_id)
db.set('experiments_admin', exp_uid, 'start_date',
utils.datetime2str(utils.datetimeNow()))
log_entry = {
'exp_uid': exp_uid,
'task': 'initExp',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.initExp input args_json is in improper format" % self.app_id
return '{}', False, error
# check for the fields that must be contained in args or error occurs
necessary_fields = ['n', 'failure_probability']
for field in necessary_fields:
try:
args_dict[field]
except KeyError:
error = "%s.initExp input arguments missing field: %s" % (
self.app_id, str(field))
return '{}', False, error
n = args_dict['n']
delta = args_dict['failure_probability']
if 'alg_list' in args_dict:
alg_list = args_dict['alg_list']
supportedAlgs = utils.get_app_supported_algs(self.app_id)
for algorithm in alg_list:
if algorithm['alg_id'] not in supportedAlgs:
error = "%s.initExp unsupported algorithm '%s' in alg_list" % (
self.app_id, alg_id)
return '{}', False, error
else:
alg_list = utils.get_app_default_alg_list(self.app_id)
if 'instructions' not in args_dict:
instructions = utils.get_app_default_instructions(app_id)
else:
instructions = args_dict['instructions']
if 'debrief' not in args_dict:
debrief = utils.get_app_default_instructions(app_id)
else:
debrief = args_dict['debrief']
if 'num_tries' not in args_dict:
num_tries = utils.get_app_default_num_tries(app_id)
else:
num_tries = args_dict['num_tries']
if 'context_type' not in args_dict:
context_type = 'none'
else:
context_type = args_dict['context_type']
if 'context' not in args_dict:
context = ''
else:
context = args_dict['context']
# ALGORITHM_MANAGEMENT_MODE FORMATTING CHECK
if 'algorithm_management_settings' not in args_dict:
params = {}
params['proportions'] = []
for algorithm in alg_list:
params['proportions'].append({
'alg_label':
algorithm['alg_label'],
'proportion':
1. / len(alg_list)
})
algorithm_management_settings = {}
algorithm_management_settings['mode'] = 'fixed_proportions'
algorithm_management_settings['params'] = params
else:
algorithm_management_settings = args_dict[
'algorithm_management_settings']
try:
mode = algorithm_management_settings['mode']
params = algorithm_management_settings['params']
except:
error = "%s.initExp algorithm_management_settings must be a dictionary with fields 'mode' and 'params'" % (
self.app_id)
return '{}', False, error
if mode == 'fixed_proportions':
try:
algorithm_proportions_list = params['proportions']
except:
error = "%s.initExp algorithm_management_settings['params'] must be a dictionary with field 'proportions'" % (
self.app_id)
return '{}', False, error
# check if alg_labels are properly labeled
for proportion_item in algorithm_proportions_list:
proportion = proportion_item['proportion']
target_alg_label = proportion_item['alg_label']
target_alg_label_in_alg_list = False
for algorithm in alg_list:
if algorithm['alg_label'] == target_alg_label:
target_alg_label_in_alg_list = True
if not target_alg_label_in_alg_list:
error = "%s.initExp algorithm_management_settings['params']['proportions'] must be a list of dictionaries, each dictionary containing the fields 'alg_label' and 'proportion'. The 'alg_label' value must be one of the alg_labels in a provided alg_list and 'proportion' must be nonnegative and sum to 1 : '%s' not in provided alg_list" % (
self.app_id, target_alg_label)
return '{}', False, error
elif mode == 'pure_exploration':
error = "%s.initExp Sorry, '%s' is not yet supported." % (
self.app_id, mode)
return '{}', False, error
elif mode == 'explore_exploit':
error = "%s.initExp Sorry, '%s' is not yet supported." % (
self.app_id, mode)
return '{}', False, error
else:
error = "%s.initExp unsupported algorithm_management_mode: '%s'. Must be in {'pure_exploration','explore_exploit','fixed_proportions'}" % (
self.app_id, algorithm_management_mode)
return '{}', False, error
# ALGORITHM_MANAGEMENT_MODE FORMATTING CHECK
if 'participant_to_algorithm_management' not in args_dict:
participant_to_algorithm_management = 'one_to_many'
else:
participant_to_algorithm_management = args_dict[
'participant_to_algorithm_management']
if participant_to_algorithm_management not in [
'one_to_many', 'one_to_one'
]:
error = "%s.initExp unsupported participant_to_algorithm_management: '%s'. Must be in {'one_to_many','one_to_one'}" % (
self.app_id, participant_to_algorithm_management)
return '{}', False, error
# assign uid to each algorithm and save it
for algorithm in alg_list:
alg_uid = utils.getNewUID()
algorithm['alg_uid'] = alg_uid
db.set(app_id + ':algorithms', alg_uid, 'alg_uid', alg_uid)
db.set(app_id + ':algorithms', alg_uid, 'exp_uid', exp_uid)
db.set(app_id + ':experiments', exp_uid, 'exp_uid', exp_uid)
db.set(app_id + ':experiments', exp_uid, 'app_id', app_id)
db.set(app_id + ':experiments', exp_uid, 'n', n)
db.set(app_id + ':experiments', exp_uid, 'failure_probability',
delta)
db.set(app_id + ':experiments', exp_uid, 'alg_list', alg_list)
db.set(app_id + ':experiments', exp_uid,
'algorithm_management_settings',
algorithm_management_settings)
db.set(app_id + ':experiments', exp_uid,
'participant_to_algorithm_management',
participant_to_algorithm_management)
db.set(app_id + ':experiments', exp_uid, 'instructions',
instructions)
db.set(app_id + ':experiments', exp_uid, 'debrief', debrief)
db.set(app_id + ':experiments', exp_uid, 'context_type',
context_type)
db.set(app_id + ':experiments', exp_uid, 'context', context)
db.set(app_id + ':experiments', exp_uid, 'num_tries', num_tries)
# now create intitialize each algorithm
for algorithm in alg_list:
alg_id = algorithm['alg_id']
alg_uid = algorithm['alg_uid']
db.set(app_id + ':algorithms', alg_uid, 'alg_id', alg_id)
db.set(app_id + ':algorithms', alg_uid, 'alg_uid', alg_uid)
db.set(app_id + ':algorithms', alg_uid, 'exp_uid', exp_uid)
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
# call initExp
didSucceed, dt = utils.timeit(alg.initExp)(
resource=rc, n=n, failure_probability=delta)
log_entry = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'initExp',
'duration': dt,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':ALG-DURATION', log_entry)
response_json = '{}'
log_entry = {
'exp_uid': exp_uid,
'task': 'initExp',
'json': response_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return response_json, True, ''
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'initExp',
'error': error,
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error
def getQuery(self, exp_uid, args_json, db, ell):
"""
A request to ask which k arms to duel next
Expected input (in jsonstructure with string keys):
(int) k : number of objects to display
[optional] (string) participant_uid : unique identifier of session for a participant answering questions (that is, an email address is not good enough as the participant could participate in multiple exp_uids so it would not be unique against all experiments), if key non-existant particpant_uid is assigned as exp_uid.
Expected output (in json structure with string keys):
(list) target_indices : list of k target indexes e.g. [ (int) target_index_1, ... , (int) target_index_k ]
(str) query_uid : unique identifier of query (used to look up for processAnswer)
"""
try:
app_id = self.app_id
log_entry = {
'exp_uid': exp_uid,
'task': 'getQuery',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.initExp input args_json is in improper format" % self.app_id
return '{}', False, error
# get list of algorithms associated with project
alg_list, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'alg_list')
alg_label_to_alg_id = {}
alg_label_to_alg_uid = {}
for algorithm in alg_list:
alg_label_to_alg_id[
algorithm['alg_label']] = algorithm['alg_id']
alg_label_to_alg_uid[
algorithm['alg_label']] = algorithm['alg_uid']
algorithm_management_settings, didSucceed, message = db.get(
app_id + ':experiments', exp_uid,
'algorithm_management_settings')
# ASSIGN ALGORITHM TO PARTICIPANT
if 'participant_uid' in args_dict:
participant_uid = args_dict['participant_uid']
else:
participant_uid = exp_uid
participant_doc_exists, didSucceed, message = db.exists(
app_id + ':participants', participant_uid, 'participant_uid')
first_participant_query = not participant_doc_exists
if first_participant_query:
db.set(app_id + ':participants', participant_uid,
'participant_uid', participant_uid)
db.set(app_id + ':participants', participant_uid, 'exp_uid',
exp_uid)
participant_to_algorithm_management, didSucceed, message = db.get(
app_id + ':experiments', exp_uid,
'participant_to_algorithm_management')
if (participant_uid == exp_uid) or (
participant_to_algorithm_management
== 'one_to_many') or (first_participant_query):
if algorithm_management_settings[
'mode'] == 'fixed_proportions':
proportions_list = algorithm_management_settings['params'][
'proportions']
prop = [
prop_item['proportion']
for prop_item in proportions_list
]
prop_item = numpy.random.choice(alg_list, p=prop)
else:
raise Exception('algorithm_management_mode : ' +
algorithm_management_settings['mode'] +
' not implemented')
alg_id = alg_label_to_alg_id[prop_item['alg_label']]
alg_uid = alg_label_to_alg_uid[prop_item['alg_label']]
alg_label = prop_item['alg_label']
if (first_participant_query) and (
participant_to_algorithm_management == 'one_to_one'):
db.set(app_id + ':participants', participant_uid, 'alg_id',
alg_id)
db.set(app_id + ':participants', participant_uid,
'alg_uid', alg_uid)
elif (participant_to_algorithm_management == 'one_to_one'):
# If here, then alg_uid should already be assigned in participant doc
alg_id, didSucceed, message = db.get(app_id + ':participants',
participant_uid, 'alg_id')
alg_uid, didSucceed, message = db.get(app_id + ':participants',
participant_uid,
'alg_uid')
else:
raise Exception('participant_to_algorithm_management : ' +
participant_to_algorithm_management +
' not implemented')
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
# call getQuery
targets, dt = utils.timeit(alg.getQuery)(resource=rc)
# update targets
target_list = []
for target in targets:
target_list.append({'index': target})
# check for context
context_type, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'context_type')
context, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'context')
# log
log_entry_durations = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'getQuery',
'duration': dt
}
log_entry_durations.update(rc.getDurations())
meta = {'log_entry_durations': log_entry_durations}
# create JSON query payload
timestamp = str(utils.datetimeNow())
query_uid = utils.getNewUID()
query = {}
query['query_uid'] = query_uid
query['target_indices'] = target_list
# save query data to database
query_doc = {}
query_doc.update(query)
query_doc['participant_uid'] = participant_uid
query_doc['alg_uid'] = alg_uid
query_doc['exp_uid'] = exp_uid
query_doc['alg_label'] = alg_label
query_doc['timestamp_query_generated'] = timestamp
for field in query_doc:
db.set(app_id + ':queries', query_uid, field, query_doc[field])
# add context after updating query doc to avoid redundant information
query['context_type'] = context_type
query['context'] = context
args_out = {'args': query, 'meta': meta}
response_json = json.dumps(args_out)
log_entry = {
'exp_uid': exp_uid,
'task': 'getQuery',
'json': response_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return response_json, True, ''
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'getQuery',
'error': error,
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error
def predict(self, exp_uid, args_json, db, ell):
"""
Description: uses current model empirical estimates to forecast the ranking of the arms in order of likelhood of being the best from most to least likely
Expected input:
(string) predict_id : 'predict_labels'
(dict) params : dictionary with fields
(string) alg_label : describes target algorithm to use
Expected output (in json structure):
(list of dicts with fields):
(int) index : index of target
(int) rank : rank that the algorithm assigns to index (rank 0 is most likely the best arm)
"""
try:
app_id = self.app_id
log_entry = {
'exp_uid': exp_uid,
'task': 'predict',
'json': args_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-CALL', log_entry)
# convert args_json to args_dict
try:
args_dict = json.loads(args_json)
except:
error = "%s.predict failed to convert input args_json due to improper format" % (
self.app_id)
return '{}', False, error
# check for the fields that must be contained in args or error occurs
necessary_fields = ['predict_id', 'params']
for field in necessary_fields:
try:
args_dict[field]
except KeyError:
error = "%s.predict input arguments missing field: %s" % (
self.app_id, str(field))
return '{}', False, error
predict_id = args_dict['predict_id']
params = args_dict['params']
if predict_id == "predict_labels":
alg_label = params['alg_label']
# get list of algorithms associated with project
alg_list, didSucceed, message = db.get(app_id + ':experiments',
exp_uid, 'alg_list')
# get alg_id
for algorithm in alg_list:
if alg_label == algorithm['alg_label']:
alg_id = algorithm['alg_id']
alg_uid = algorithm['alg_uid']
num_reported_answers, didSucceed, message = db.get(
app_id + ':experiments', exp_uid,
'num_reported_answers_for_' + alg_uid)
if type(num_reported_answers) != int:
num_reported_answers = 0
# get sandboxed database for the specific app_id,alg_id,exp_uid - closing off the rest of the database to the algorithm
rc = ResourceClient(app_id, exp_uid, alg_uid, db)
# get specific algorithm to make calls to
alg = utils.get_app_alg(self.app_id, alg_id)
# call getQuery
predicted_labels, dt = utils.timeit(alg.predict)(resource=rc)
log_entry_durations = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'task': 'predict',
'duration': dt
}
log_entry_durations.update(rc.getDurations())
meta = {'log_entry_durations': log_entry_durations}
test_labels, didSucceed, message = db.get(
app_id + ':experiments', exp_uid, 'test_labels')
num_errors = 0
for answer in test_labels:
target_index, target_label = answer
if target_label * predicted_labels[target_index] < 0:
num_errors += 1.
error = num_errors / float(max(1., len(test_labels)))
log_entry = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'timestamp': utils.datetimeNow()
}
log_entry.update({
'error': error,
'num_reported_answers': num_reported_answers
})
ell.log(app_id + ':ALG-EVALUATION', log_entry)
response_args_dict = {
'exp_uid': exp_uid,
'alg_uid': alg_uid,
'error': error,
'num_reported_answers': num_reported_answers
}
args_out = {'args': response_args_dict, 'meta': meta}
predict_json = json.dumps(args_out)
log_entry = {
'exp_uid': exp_uid,
'task': 'predict',
'json': predict_json,
'timestamp': utils.datetimeNow()
}
ell.log(app_id + ':APP-RESPONSE', log_entry)
return predict_json, True, ''
except Exception, err:
error = traceback.format_exc()
log_entry = {
'exp_uid': exp_uid,
'task': 'predict',
'error': str(error),
'timestamp': utils.datetimeNow(),
'args_json': args_json
}
didSucceed, message = ell.log(app_id + ':APP-EXCEPTION', log_entry)
return '{}', False, error