def mHealth_rollout(tuning_function_parameter, policy, time_horizon, estimated_context_mean,
tuning_function, estimated_context_variance, env, nPatients, monte_carlo_reps):
score = 0
rollout_env = NormalCB(list_of_reward_betas=env.beta_hat_list, list_of_reward_vars=env.sigma_hat_list,
context_mean=estimated_context_mean, context_var=estimated_context_variance)
for rep in range(monte_carlo_reps):
rollout_env.reset()
episode_score = 0
# Initial assignments
for t in range(10):
for j in range(5):
rollout_env.step(0)
for j in range(5):
rollout_env.step(1)
for time in range(time_horizon):
beta_hat = rollout_env.beta_hat_list
sampling_cov_list = rollout_env.sampling_cov_list
for j in range(nPatients):
# Draw context and take action
# context = context_sequence[time - current_time][j]
action = policy(beta_hat, sampling_cov_list, rollout_env.curr_context, tuning_function,
tuning_function_parameter, time_horizon, time, env)
expected_reward = rollout_env.expected_reward(action, rollout_env.curr_context)
optimal_expected_reward = np.max([rollout_env.expected_reward(a, rollout_env.curr_context)
for a in range(rollout_env.number_of_actions)])
rollout_env.step(action)
# Update regret
regret = (expected_reward - optimal_expected_reward)
episode_score += regret
print(rep)
score += (episode_score - score) / (rep + 1)
return score
def episode(policy_name,
label,
save=False,
points_per_grid_dimension=50,
monte_carlo_reps=100):
if save:
base_name = 'mhealth-{}-{}'.format(label, policy_name)
prefix = os.path.join(project_dir, 'src', 'run', 'results', base_name)
suffix = datetime.datetime.now().strftime("%y%m%d_%H%M%S")
filename = '{}_{}.yml'.format(prefix, suffix)
np.random.seed(label)
T = 10
# ToDo: Create policy class that encapsulates this behavior
if policy_name == 'eps':
tuning_function = lambda a, b, c: 0.05 # Constant epsilon
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'eps-decay-fixed':
tuning_function = lambda a, t, c: 0.5 / (t + 1)
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'eps-decay':
tuning_function = tuned_bandit.stepwise_linear_epsilon
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = True
tuning_function_parameter = np.ones(10) * 0.025
elif policy_name == 'greedy':
tuning_function = lambda a, b, c: 0.00 # Constant epsilon
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'worst':
tuning_function = lambda a, b, c: 0.00
policy = ref.linear_cb_worst_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'ts':
tuning_function = lambda a, b, c: 1.0 # No shrinkage
policy = tuned_bandit.linear_cb_thompson_sampling_policy
tune = False
tuning_function_parameter = None
# elif policy_name == 'ts-shrink':
# tuning_function = tuned_bandit.expit_truncate
# policy = tuned_bandit.thompson_sampling_policy
# tune = True
# tuning_function_parameter = np.array([-2, 1])
else:
raise ValueError('Incorrect policy name')
env = NormalCB(
list_of_reward_betas=[np.array([1.0, 1.0]),
np.array([2.0, -2.0])])
cumulative_regret = 0.0
nPatients = 10
env.reset()
# Initial assignments
for t in range(10):
for j in range(5):
env.step(0)
for j in range(5):
env.step(1)
for t in range(T):
X = env.X
estimated_context_mean = np.mean(X, axis=0)
estimated_context_variance = np.cov(X, rowvar=False)
if tune:
tuning_function_parameter = opt.bayesopt(
rollout.mHealth_rollout, policy, tuning_function,
tuning_function_parameter, T, estimated_context_mean,
estimated_context_variance, env, nPatients,
points_per_grid_dimension, monte_carlo_reps)
# print('time {} epsilon {}'.format(t, tuning_function(T,t,tuning_function_parameter)))
for j in range(nPatients):
x = copy.copy(env.curr_context)
beta_hat = env.beta_hat_list
action = policy(beta_hat, env.sampling_cov_list, x,
tuning_function, tuning_function_parameter, T, t,
env)
env.step(action)
# Compute regret
expected_rewards = [
env.expected_reward(a, env.curr_context)
for a in range(env.number_of_actions)
]
expected_reward_at_action = expected_rewards[action]
optimal_expected_reward = np.max(expected_rewards)
regret = optimal_expected_reward - expected_reward_at_action
cumulative_regret += regret
# Save results
if save:
results = {'t': float(t), 'regret': float(cumulative_regret)}
with open(filename, 'w') as outfile:
yaml.dump(results, outfile)
return cumulative_regret
def episode(policy_name,
label,
n_patients=15,
list_of_reward_betas=[[-10, 0.4, 0.4, -0.4],
[-9.8, 0.6, 0.6, -0.4]],
context_mean=np.array([0.0, 0.0, 0.0]),
context_var=np.array([[1.0, 0, 0], [0, 1., 0], [0, 0, 1.]]),
list_of_reward_vars=[1, 1],
T=50,
mc_replicates=1000,
pre_simulate=True):
np.random.seed(label)
# ToDo: Create policy class that encapsulates this behavior
posterior_sample = True
bootstrap_posterior = False
positive_zeta = False
if policy_name == 'eps':
tuning_function = lambda a, b, c: 0.1 # Constant epsilon
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'random':
tuning_function = lambda a, b, c: 1.0 # Constant epsilon
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'eps-decay-fixed':
tuning_function = tuned_bandit.expit_epsilon_decay
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = np.array([0.8, 46.38, 1.857])
elif policy_name == 'eps-decay':
tuning_function = tuned_bandit.expit_epsilon_decay
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = True
explore_ = {
'zeta0': [1.0, 0.05, 1.0, 0.1],
'zeta1': [30.0, 0.0, 1.0, 0.0],
'zeta2': [0.1, 1.0, 0.01, 1.0]
}
bounds = {
'zeta0': (0.025, 2.0),
'zeta1': (0.0, 30.0),
'zeta2': (0.01, 2)
}
tuning_function_parameter = np.array([0.05, 1.0, 0.01])
posterior_sample = True
elif policy_name == 'greedy':
tuning_function = lambda a, b, c: 0.00 # Constant epsilon
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'worst':
tuning_function = lambda a, b, c: 0.00
policy = ref.linear_cb_worst_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'ts':
tuning_function = lambda a, b, c: 1.0 # No shrinkage
policy = tuned_bandit.linear_cb_thompson_sampling_policy
tune = False
tuning_function_parameter = None
elif policy_name == 'ts-decay-posterior-sample':
tuning_function = tuned_bandit.stepwise_linear_epsilon
policy = tuned_bandit.linear_cb_thompson_sampling_policy
tune = True
tuning_function_parameter = np.ones(10) * 0.1
posterior_sample = True
elif policy_name == 'ts-decay-bootstrap-sample':
tuning_function = tuned_bandit.stepwise_linear_epsilon
policy = tuned_bandit.linear_cb_thompson_sampling_policy
tune = True
tuning_function_parameter = np.ones(10) * 0.1
posterior_sample = True
bootstrap_posterior = True
elif policy_name == 'ts-decay':
tuning_function = tuned_bandit.stepwise_linear_epsilon
policy = tuned_bandit.linear_cb_thompson_sampling_policy
tune = True
tuning_function_parameter = np.ones(10) * 0.1
elif policy_name == 'ucb-tune-posterior-sample':
tuning_function = tuned_bandit.stepwise_linear_epsilon
policy = tuned_bandit.linear_cb_ucb_policy
tune = True
tuning_function_parameter = np.ones(10) * 0.025
posterior_sample = True
# elif policy_name == 'ts-shrink':
# tuning_function = tuned_bandit.expit_truncate
# policy = tuned_bandit.thompson_sampling_policy
# tune = True
# tuning_function_parameter = np.array([-2, 1])
else:
raise ValueError('Incorrect policy name')
env = NormalCB(list_of_reward_betas=list_of_reward_betas,
context_mean=context_mean,
context_var=context_var,
list_of_reward_vars=list_of_reward_vars)
# env = NormalUniformCB(list_of_reward_betas=[np.ones(10) + 0.05, np.ones(10)], list_of_reward_vars=[0.01, 25])
cumulative_regret = 0.0
# env.reset()
tuning_parameter_sequence = []
rewards = []
actions = []
# Using pre-simulated data
# data_for_episode = env.generate_mc_samples(1, T)
# rep_dict = data_for_episode[0]
# initial_linear_model = rep_dict['initial_linear_model']
# beta_hat_list = initial_linear_model['beta_hat_list']
# Xprime_X_list = initial_linear_model['Xprime_X_list']
# Xprime_X_inv_list = initial_linear_model['Xprime_X_inv_list']
# X_list = initial_linear_model['X_list']
# y_list = initial_linear_model['y_list']
# X_dot_y_list = initial_linear_model['X_dot_y_list']
# sampling_cov_list = initial_linear_model['sampling_cov_list']
# sigma_hat_list = initial_linear_model['sigma_hat_list']
# context_sequence = rep_dict['contexts']
# regrets_sequence = rep_dict['regrets']
# rewards_sequence = rep_dict['rewards']
for t in range(T):
X = env.X
estimated_context_mean = np.mean(X, axis=0)
estimated_context_variance = np.cov(X, rowvar=False)
estimated_context_bounds = (np.min(X), np.max(X[:, 1:]))
if tune:
if pre_simulate:
if posterior_sample:
gen_model_parameters = []
for rep in range(mc_replicates):
if bootstrap_posterior:
pass
else:
draws = env.sample_from_posterior()
# draws = env.sample_from_sampling_dist()
betas_for_each_action = []
vars_for_each_action = []
for a in range(env.number_of_actions):
beta_a = draws[a]['beta_draw']
var_a = draws[a]['var_draw']
betas_for_each_action.append(beta_a)
vars_for_each_action.append(var_a)
param_dict = {
'reward_betas': betas_for_each_action,
'reward_vars': vars_for_each_action,
'context_mean': draws['context_mu_draw'],
'context_var': draws['context_var_draw']
}
# 'context_max': draws['context_max']}
gen_model_parameters.append(param_dict)
else:
gen_model_parameters = None
# sim_env = NormalUniformCB(list_of_reward_betas=env.beta_hat_list, list_of_reward_vars=env.sigma_hat_list,
# context_bounds=estimated_context_bounds)
sim_env = NormalCB(list_of_reward_betas=list_of_reward_betas,
context_mean=context_mean,
context_var=context_var,
list_of_reward_vars=list_of_reward_vars)
pre_simulated_data = sim_env.generate_mc_samples(
mc_replicates,
T,
n_patients=n_patients,
gen_model_params=gen_model_parameters)
tuning_function_parameter = opt.bayesopt(
rollout.normal_cb_rollout_with_fixed_simulations,
policy,
tuning_function,
tuning_function_parameter,
T,
sim_env,
mc_replicates, {'pre_simulated_data': pre_simulated_data},
bounds,
explore_,
positive_zeta=positive_zeta)
tuning_parameter_sequence.append(
[float(z) for z in tuning_function_parameter])
else:
tuning_function_parameter = tuned_bandit.random_search(
tuned_bandit.oracle_rollout, policy, tuning_function,
tuning_function_parameter, linear_model_results, T, t,
estimated_context_mean, estimated_context_variance, env)
for patient in range(n_patients):
x = copy.copy(env.curr_context)
beta_hat = np.array([
env.posterior_params_dict[a]['beta_post']
for a in range(env.number_of_actions)
])
# print(env.posterior_params_dict)
action = policy(beta_hat, env.sampling_cov_list, x,
tuning_function, tuning_function_parameter, T, t,
env)
res = env.step(action)
cumulative_regret += -env.regret(action, x)
actions.append(action)
u = res['Utility']
rewards.append(u)
print(beta_hat)
if t == 0:
break
return {
'cumulative_regret': cumulative_regret,
'zeta_sequence': tuning_parameter_sequence,
'rewards': rewards,
'actions': actions
}
def episode(label,
tuning_function_parameter,
n_patients=1,
list_of_reward_betas=[[-10, 0.4, 0.4, -0.4],
[-9.8, 0.6, 0.6, -0.4]],
context_mean=np.array([0.0, 0.0, 0.0]),
context_var=np.array([[1.0, 0, 0], [0, 1., 0], [0, 0, 1.]]),
list_of_reward_vars=[1, 1],
T=30):
tuning_function = tuned_bandit.expit_epsilon_decay
policy = tuned_bandit.linear_cb_epsilon_greedy_policy
env = NormalCB(1,
list_of_reward_betas=list_of_reward_betas,
context_mean=context_mean,
context_var=context_var,
list_of_reward_vars=list_of_reward_vars)
# env = NormalUniformCB(list_of_reward_betas=[np.ones(10) + 0.05, np.ones(10)], list_of_reward_vars=[0.01, 25])
cumulative_regret = 0.0
# env.reset()
print('epsilon', tuning_function(T, 0, tuning_function_parameter))
tuning_parameter_sequence = []
rewards = []
actions = []
# Using pre-simulated data
# data_for_episode = env.generate_mc_samples(1, T)
# rep_dict = data_for_episode[0]
# initial_linear_model = rep_dict['initial_linear_model']
# beta_hat_list = initial_linear_model['beta_hat_list']
# Xprime_X_list = initial_linear_model['Xprime_X_list']
# Xprime_X_inv_list = initial_linear_model['Xprime_X_inv_list']
# X_list = initial_linear_model['X_list']
# y_list = initial_linear_model['y_list']
# X_dot_y_list = initial_linear_model['X_dot_y_list']
# sampling_cov_list = initial_linear_model['sampling_cov_list']
# sigma_hat_list = initial_linear_model['sigma_hat_list']
# context_sequence = rep_dict['contexts']
# regrets_sequence = rep_dict['regrets']
# rewards_sequence = rep_dict['rewards']
for t in range(T):
X = env.X
for patient in range(n_patients):
x = copy.copy(env.curr_context)
beta_hat = np.array([
env.posterior_params_dict[a]['beta_post']
for a in range(env.number_of_actions)
])
# print(env.posterior_params_dict)
action = policy(beta_hat, env.sampling_cov_list, x,
tuning_function, tuning_function_parameter, T, t,
env)
res = env.step(action)
cumulative_regret += -env.regret(action, x)
actions.append(int(action))
u = res['Utility']
rewards.append(float(u))
return {
'cumulative_regret': float(cumulative_regret),
'rewards': rewards,
'actions': actions
}