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 main(args):
scenario = EnvironmentManager.load_scenario(args.scenario_name)
env = PricingAdvertisingJointEnvironment(scenario)
env.set_budget_allocation([args.budget] * scenario.get_n_subcampaigns())
prices = get_prices(args=args)
arm_profit = prices - args.unit_cost
bandit = build_discrete_bandit(bandit_name=args.bandit_name, n_arms=len(arm_profit),
arm_values=arm_profit, args=args)
iterate = not env.next_day()
while iterate:
iterate = not env.next_day()
for _ in range(0, args.n_rounds):
# Choose arm
price_idx = bandit.pull_arm()
# Observe reward
env.next_user()
reward, elapsed_day = env.round(price=prices[price_idx])
reward = reward * arm_profit[price_idx]
if elapsed_day:
iterate = not env.next_day()
while iterate:
iterate = not env.next_day()
# Update bandit
bandit.update(pulled_arm=price_idx, reward=reward)
return bandit.collected_rewards, env.get_day_breakpoints()
def main(args):
# Retrieve scenario
scenario = EnvironmentManager.load_scenario(args.scenario_name)
# Retrieve bandit and basic information
campaign = Campaign(scenario.get_n_subcampaigns(), args.cum_budget,
args.n_arms_ads)
prices = get_prices(args=args)
arm_profit = prices - args.unit_cost
bandit = get_bandit(args=args, arm_values=arm_profit, campaign=campaign)
# Create environment
env = PricingAdvertisingJointEnvironment(scenario=scenario)
# choose whether to pull one arm a day for each class, or to pull one arm for each user
if args.daily_price:
learn_per_day(bandit, env, campaign, prices, arm_profit,
scenario.get_n_subcampaigns())
else:
learn_per_user(bandit, env, campaign, prices, arm_profit)
return bandit.get_daily_reward(), env.day_breakpoints
inplace=True)
mean_df.rename(columns={n_bandit: "day"}, inplace=True)
for bandit_idx, name in enumerate(BANDIT_NAME):
std_df.rename(columns={bandit_idx: "mean_std_{}".format(name)},
inplace=True)
std_df.rename(columns={n_bandit: "day"}, inplace=True)
total_df = mean_df.merge(std_df, left_on="day", right_on="day")
total_df.to_csv("{}instant_reward.csv".format(folder_path_with_date),
index=False)
# Something
mean_scenario: Scenario = EnvironmentManager.load_scenario(
SCENARIO_NAME, get_mean_function=True)
click_function_list: List[IStochasticFunction] = mean_scenario.get_phases(
)[0].get_n_clicks_function()
# Optimal point computation
campaign = Campaign(n_sub_campaigns=mean_scenario.get_n_subcampaigns(),
cum_budget=BUDGET,
n_arms=N_ARMS_ADV)
for i in range(campaign.get_n_sub_campaigns()):
sub_campaign_values = [
click_function_list[i].draw_sample(b)
for b in np.linspace(0, BUDGET, N_ARMS_ADV)
]
campaign.set_sub_campaign_values(i, sub_campaign_values)
max_clicks, best_budgets = CampaignOptimizer.find_best_budgets(campaign)
args.pricing_bandit_name))
fd.write("Bandit algorithm for ads: {}\n".format(args.ads_bandit_name))
fd.write("User prices: {}\n".format(get_prices(args=args)))
fd.write("Cumulative budget: {}\n\n".format(args.cum_budget))
fd.write("Bandit parameters \n")
fd.write(
"Min quantile standard deviation (For quantile bandits): {}\n".format(
args.min_std_q))
fd.write("Gamma parameter (EXP3) {}\n".format(args.gamma))
fd.write("CRP upper bound {}\n".format(args.crp_upper_bound))
fd.write("Alpha GP: {}\n".format(args.alpha))
fd.write("Number of GP optimizer restarts: {}\n".format(
args.n_restart_opt))
fd.write("Initial standard deviation GP: {}\n".format(args.init_std))
fd.close()
# Plot instantaneous reward
rewards = np.mean(total_rewards, axis=0)
scenario = EnvironmentManager.load_scenario(args.scenario_name)
env = PricingAdvertisingJointEnvironment(scenario)
os.chdir(folder_path_with_date)
plt.plot(rewards, 'g')
plt.xlabel("t")
plt.ylabel("Instantaneous Reward")
plt.title(str(args.n_runs) + " Experiments")
plt.savefig(fname="Reward.png", format="png")