def _optimize(cls, campaign: Campaign) -> (np.ndarray, np.ndarray):
"""
Optimize the combinatorial problem of the advertising campaign by using a dynamic programming algorithm
:param campaign: the campaign to be optimized
:return:
- the optimization matrix (N+1) x M containing, for each pair (budget, set of sub-campaign), the maximum
optimized value achievable (value might be, for instance, the number of clicks, or the profit)
- the maximum indices (N+1) x M related to the optimization matrix containing, for each pair
(budget, set of sub-campaign), the index of the best budget for the new added sub-campaign w.r.t. previous
set of sub-campaign (i.e. row)
where N is the number of sub-campaign and M is the number of discrete budgets
"""
optimization_matrix = np.zeros(shape=(campaign.get_n_sub_campaigns() +
1, len(campaign.get_budgets())))
max_idx_matrix = np.full_like(optimization_matrix,
fill_value=-1,
dtype=np.int)
prev_row = 0
clicks_matrix = campaign.get_sub_campaigns()
for row in range(1, optimization_matrix.shape[0]):
temp_clicks = clicks_matrix[row - 1][::-1]
for col in range(optimization_matrix.shape[1]):
cum_sum_clicks = temp_clicks[optimization_matrix.shape[1] -
col - 1:] + optimization_matrix[
prev_row, :col + 1]
idx_max = np.argmax(cum_sum_clicks)
optimization_matrix[row, col] = cum_sum_clicks[idx_max]
max_idx_matrix[row, col] = col - idx_max
prev_row = row
return optimization_matrix, max_idx_matrix
def main(args):
scenario = EnvironmentManager.load_scenario(args.scenario_name)
env = PricingAdvertisingJointEnvironment(scenario)
campaign = Campaign(scenario.get_n_subcampaigns(), args.cum_budget,
args.n_arms)
bandit = build_combinatorial_bandit(bandit_name=args.bandit_name,
campaign=campaign,
init_std=args.init_std,
args=args)
budget_allocation = [0, 0, 0]
for t in range(0, args.n_rounds):
# Choose arm
budget_allocation_indexes = bandit.pull_arm()
budget_allocation = [
int(campaign.get_budgets()[i]) for i in budget_allocation_indexes
]
# Observe reward
env.set_budget_allocation(budget_allocation=budget_allocation)
env.next_day()
rewards = env.get_daily_visits_per_sub_campaign()
# Update bandit
bandit.update(pulled_arm=budget_allocation_indexes,
observed_reward=rewards)
return bandit.collected_rewards, budget_allocation
def get_bandit(args, arm_values: np.array, campaign: Campaign) -> IJointBandit:
bandit_name: str = args.joint_bandit_name
ads_bandit = build_combinatorial_bandit(bandit_name=args.ads_bandit_name,
campaign=campaign,
init_std=args.init_std,
args=args)
price_bandit_class, price_bandit_kwargs = get_bandit_class_and_kwargs(
bandit_name=args.pricing_bandit_name,
n_arms=len(arm_values),
arm_values=arm_values,
args=args)
price_bandit_list = [
price_bandit_class(**price_bandit_kwargs)
for _ in range(campaign.get_n_sub_campaigns())
]
ad_value_strategy = ExpectationAdValueStrategy(np.max(arm_values)) \
if bandit_name.find("Exp") >= 0 else QuantileAdValueStrategy(np.max(arm_values), args.min_std_q)
is_learn_visits = True if bandit_name[-1] == 'V' else False
if bandit_name in ["JBExp", "JBExpV", "JBQV", "JBQ"]:
bandit = JointBanditDiscriminatory(ads_learner=ads_bandit,
price_learner=price_bandit_list,
campaign=campaign,
ad_value_strategy=ad_value_strategy,
is_learn_visits=is_learn_visits)
elif bandit_name in ["JBIExpV", "JBIQV"]:
bandit = JointBanditDiscriminatoryImproved(
ads_learner=ads_bandit,
price_learner=price_bandit_list,
campaign=campaign,
ad_value_strategy=ad_value_strategy)
elif bandit_name in ["JBBQ", "JBBExp"]:
model_list: List[DiscreteRegressor] = [
DiscreteGPRegressor(list(campaign.get_budgets()),
args.init_std,
args.alpha,
args.n_restart_opt,
normalized=True)
for _ in range(campaign.get_n_sub_campaigns())
]
bandit = JointBanditBalanced(campaign=campaign,
arm_values=arm_values,
price_learner_class=price_bandit_class,
price_learner_kwargs=price_bandit_kwargs,
number_of_visit_model_list=model_list,
ad_value_strategy=ad_value_strategy)
elif bandit_name == "JBFQ":
assert args.daily_price, "This joint bandit requires to run in a daily manner"
model_list: List[DiscreteRegressor] = [
DiscreteGPRegressor(list(campaign.get_budgets()),
args.init_std,
args.alpha,
args.n_restart_opt,
normalized=True)
for _ in range(campaign.get_n_sub_campaigns())
]
bandit = JointBanditFixedDailyPriceQuantile(
campaign=campaign,
number_of_visit_model_list=model_list,
min_std=args.min_std_q,
arm_profit=arm_values,
n_arms_profit=len(arm_values))
elif bandit_name == "JBFTS":
assert args.daily_price, "This joint bandit requires to run in a daily manner"
model_list: List[DiscreteRegressor] = [
DiscreteGPRegressor(list(campaign.get_budgets()),
args.init_std,
args.alpha,
args.n_restart_opt,
normalized=True)
for _ in range(campaign.get_n_sub_campaigns())
]
bandit = JointBanditFixedDailyPriceTS(
campaign=campaign,
number_of_visit_model_list=model_list,
arm_profit=arm_values,
n_arms_profit=len(arm_values))
else:
raise argparse.ArgumentParser(
"The name of the bandit to be used is not in the available ones")
return bandit