def runFactorialThompsonContextualBandit(config):
'''
Just for experimenting with running the contextual bandit code. This uses the factorial version, so
the actions are chosen separately for each experiment.
'''
# Initial experiments - treat this as a factorial - one choice at each timestep but lots of actions
# First, get the right models to include regression
conditions = config["conditions"]
numConditions = [len(curConditions) for curConditions in conditions]
conditionVectors = makeConditionVectors(numConditions)
conditionsToActionIndex = {
tuple(vector): i
for (i, vector) in zip(range(len(conditionVectors)), conditionVectors)
}
models = [[
logistic_regression.RLogReg(D=6, Lambda=1) for _ in range(curNum)
] for curNum in numConditions]
# models = [logistic_regression.RLogReg(D=6, Lambda=1) for _ in range(len(conditionVectors))]
# Then, run thompson sampling
get_context = lambda row: contextualStructure.get_context(
row, includeIntercept=True)
chosen_actions, models = thompson_policy.calculate_thompson_single_bandit_factorial(
config["rewardFile"],
numConditions,
config["outfilePrefix"] + "FactorialArms.csv",
models,
conditionsToActionIndex,
get_context=get_context)
for curModels in models:
print([model.w for model in curModels])
def runUniformRandomBandit(config):
conditions = config["conditions"]
numConditions = [len(curConditions) for curConditions in conditions]
conditionVectors = makeConditionVectors(numConditions)
models = [logistic_regression.RLogReg(D=6, Lambda=1) for _ in range(len(conditionVectors))]
# Then, run uniform random sampling
get_context = lambda row: contextualStructure.get_context(row, includeIntercept=True)
# epsilon is 1 below, so an action is always chosen uniformly at random
chosen_actions, models = thompson_policy.calculate_thompson_single_bandit(config["rewardFile"],
len(conditionVectors),
config["outfilePrefix"] + "Random.csv",
models,
epsilon = 1,
get_context=get_context)
def runThompsonContextualBandit(config):
'''
Just for experimenting with running the contextual bandit code.
'''
# Initial experiments - treat this as a factorial - one choice at each timestep but lots of actions
# First, get the right models to include regression
conditions = config["conditions"]
numConditions = [len(curConditions) for curConditions in conditions]
conditionVectors = makeConditionVectors(numConditions)
models = [logistic_regression.RLogReg(D=6, Lambda=1) for _ in range(len(conditionVectors))]
# Then, run thompson sampling
get_context = lambda row: contextualStructure.get_context(row, includeIntercept=True)
chosen_actions, models = thompson_policy.calculate_thompson_single_bandit(config["rewardFile"],
len(conditionVectors),
config["outfilePrefix"] + "ExpArms.csv",
models,
get_context=get_context)
print([model.w for model in models])