def run_simulations_empirical_rewards(num_sims,
reward_file,
experiment_id,
reward_header,
is_cost,
outfile_directory,
prior_mean=0,
forceActions=0,
shuffle_data=False):
'''
Runs num_sims bandit simulations with several different sample sizes (those in the list step_sizes).
Bandit uses the thompson_ng sampling policy. Assumes reward_file is formatted like ASSISTments data,
where the reward is present under the column reward_header. Runs for as many steps as it's able
to gain samples
'''
num_actions = 2
max_steps = -1
means = []
variance = []
for i in range(num_sims):
arm_1_rewards, arm_2_rewards = get_assistments_rewards.read_assistments_rewards(
reward_file, reward_header, experiment_id, is_cost)
if shuffle_data:
random.shuffle(arm_1_rewards)
random.shuffle(arm_2_rewards)
max_steps = len(arm_1_rewards) + len(arm_2_rewards)
means = [np.mean(arm_1_rewards), np.mean(arm_2_rewards)]
variance = [np.var(arm_1_rewards), np.var(arm_2_rewards)]
if forceActions != 0:
print("Forcing actions:", forceActions)
forced = make_forced_actions(
num_actions,
len(arm_1_rewards) + len(arm_2_rewards), forceActions)
else:
forced = forced_actions()
cur_output_file = get_output_filename(
outfile_directory,
len(arm_1_rewards) + len(arm_2_rewards), i)
models = [
ng_normal.NGNormal(mu=prior_mean, k=1, alpha=1, beta=1)
for _ in range(num_actions)
]
thompson_ng_policy.calculate_thompson_single_bandit_empirical_params(
arm_1_rewards,
arm_2_rewards,
num_actions=num_actions,
dest=cur_output_file,
models=models,
action_mode=thompson_ng_policy.ActionSelectionMode.prob_is_best,
relearn=True,
forced=forced)
return max_steps, means, variance
def run_simulations_empirical_rewards(num_sims,
reward_file,
experiment_id,
reward_header,
is_cost,
outfile_directory,
successPrior=1,
failurePrior=1,
forceActions=0,
shuffle_data=False):
'''
Runs num_sims bandit simulations with several different sample sizes (those in the list step_sizes).
Bandit uses the thompson_ng sampling policy.
'''
num_actions = 2
max_steps = -1
means = []
variance = []
for i in range(num_sims):
arm_1_rewards, arm_2_rewards = get_assistments_rewards.read_assistments_rewards(
reward_file, reward_header, experiment_id, is_cost)
if shuffle_data:
random.shuffle(arm_1_rewards)
random.shuffle(arm_2_rewards)
max_steps = len(arm_1_rewards) + len(arm_2_rewards)
means = [np.mean(arm_1_rewards), np.mean(arm_2_rewards)]
variance = [np.var(arm_1_rewards), np.var(arm_2_rewards)]
if forceActions != 0:
print("Forcing actions:", forceActions)
forced = run_effect_size_simulations.make_forced_actions(
num_actions,
len(arm_1_rewards) + len(arm_2_rewards), forceActions)
else:
forced = forced_actions()
cur_output_file = get_output_filename(
outfile_directory,
len(arm_1_rewards) + len(arm_2_rewards), i)
models = [
beta_bernoulli.BetaBern(success=successPrior, failure=failurePrior)
for _ in range(num_actions)
]
thompson_policy.calculate_thompson_single_bandit_empirical_params(
arm_1_rewards,
arm_2_rewards,
num_actions=num_actions,
dest=cur_output_file,
models=models,
action_mode=thompson_policy.ActionSelectionMode.prob_is_best,
relearn=True,
forced=forced)
return max_steps, means, variance