def policy_generation(bandit, actions):
historystorage = history.MemoryHistoryStorage()
modelstorage = model.MemoryModelStorage()
if bandit == 'Exp4P':
policy = exp4p.Exp4P(actions,
historystorage,
modelstorage,
delta=0.5,
pmin=None)
elif bandit == 'LinUCB':
policy = linucb.LinUCB(actions, historystorage, modelstorage, 0.3, 20)
elif bandit == 'LinThompSamp':
policy = linthompsamp.LinThompSamp(actions,
historystorage,
modelstorage,
d=20,
delta=0.61,
r=0.01,
epsilon=0.71)
elif bandit == 'UCB1':
policy = ucb1.UCB1(actions, historystorage, modelstorage)
elif bandit == 'Exp3':
policy = exp3.Exp3(actions, historystorage, modelstorage, gamma=0.2)
elif bandit == 'random':
policy = 0
return policy
def test_reward_order_descending(self):
policy = exp3.Exp3(self.actions, self.historystorage, self.modelstorage, self.gamma)
history_id, action = policy.get_action([1, 1])
history_id_2, action_2 = policy.get_action([3, 3])
policy.reward(history_id_2, 1)
self.assertTrue(
(policy._historystorage.get_history(history_id).context == np.transpose(np.array([[1, 1]]))).all())
self.assertTrue(
(policy._historystorage.get_history(history_id_2).context == np.transpose(np.array([[3, 3]]))).all())
self.assertEqual(policy._historystorage.get_history(history_id).reward, None)
self.assertEqual(policy._historystorage.get_history(history_id_2).reward, 1)
def policy_generation(bandit, actions):
"""
Parameters
----------
bandit： 赌博机算法
actions：动作即推荐的电影
Returns
-------
policy： 生成的策略
"""
historystorage = history.MemoryHistoryStorage() # 内存中历史存储记录
modelstorage = model.MemoryModelStorage() # 内存中模型存储，为了统一
if bandit == 'Exp4P':
policy = exp4p.Exp4P(historystorage,
modelstorage,
actions,
delta=0.5,
p_min=None)
elif bandit == 'LinUCB':
#policy = linucb.LinUCB(historystorage, modelstorage, actions, 0.3, 20)
policy = linucb.LinUCB(history_storage=historystorage,
model_storage=modelstorage,
action_storage=actions,
alpha=0.3,
context_dimension=18)
elif bandit == 'LinThompSamp':
policy = linthompsamp.LinThompSamp(
historystorage,
modelstorage,
actions,
#d=20, Supposed to be context dimension
context_dimension=18,
delta=0.61,
R=0.01,
epsilon=0.71)
elif bandit == 'UCB1':
policy = ucb1.UCB1(historystorage, modelstorage, actions)
elif bandit == 'Exp3':
policy = exp3.Exp3(historystorage, modelstorage, actions, gamma=0.2)
elif bandit == 'random':
policy = 0
return policy
def policy_generation(bandit, actions):
historystorage = history.MemoryHistoryStorage()
modelstorage = model.MemoryModelStorage()
if bandit == 'Exp4P':
policy = exp4p.Exp4P(historystorage,
modelstorage,
actions,
delta=0.5,
p_min=None)
elif bandit == 'LinUCB':
#policy = linucb.LinUCB(historystorage, modelstorage, actions, 0.3, 20)
policy = linucb.LinUCB(history_storage=historystorage,
model_storage=modelstorage,
action_storage=actions,
alpha=0.3,
context_dimension=18)
elif bandit == 'LinThompSamp':
policy = linthompsamp.LinThompSamp(
historystorage,
modelstorage,
actions,
#d=20, Supposed to be context dimension
context_dimension=18,
delta=0.61,
R=0.01,
epsilon=0.71)
elif bandit == 'UCB1':
policy = ucb1.UCB1(historystorage, modelstorage, actions)
elif bandit == 'Exp3':
policy = exp3.Exp3(historystorage, modelstorage, actions, gamma=0.2)
elif bandit == 'random':
policy = 0
return policy
def policy_evaluation(self, policy, context, desired_action, gamma):
if policy != 'EXP3':
print("We don't support other bandit algorithms now!")
else:
historystorage = history.MemoryHistoryStorage()
modelstorage = model.MemoryModelStorage()
seq_error = np.zeros(shape=(self.t, 1))
policy = exp3.Exp3(self.actions, historystorage, modelstorage,
gamma)
for t in range(self.t):
history_id, action = policy.get_action(context[t])
if desired_action[t][0] != action:
policy.reward(history_id, 0)
# sum_error += 1
if t == 0:
seq_error[t] = 1.0
else:
seq_error[t] = seq_error[t - 1] + 1.0
else:
policy.reward(history_id, 1)
if t > 0:
seq_error[t] = seq_error[t - 1]
return seq_error
def test_model_storage(self):
policy = exp3.Exp3(self.actions, self.historystorage, self.modelstorage, self.gamma)
history_id, action = policy.get_action([1, 1])
policy.reward(history_id, 1.0)
self.assertEqual(len(policy._modelstorage._model['w']), 5)
self.assertEqual(len(policy._modelstorage._model['query_vector']), 5)
def test_update_reward(self):
policy = exp3.Exp3(self.actions, self.historystorage, self.modelstorage, self.gamma)
history_id, action = policy.get_action([1, 1])
policy.reward(history_id, 1.0)
self.assertEqual(policy._historystorage.get_history(history_id).reward, 1.0)
def test_get_first_action(self):
policy = exp3.Exp3(self.actions, self.historystorage, self.modelstorage, self.gamma)
history_id, action = policy.get_action([1, 1])
self.assertEqual(history_id, 0)
self.assertIn(action, self.actions)
def test_initialization(self):
policy = exp3.Exp3(self.actions, self.historystorage, self.modelstorage, self.gamma)
self.assertEqual(self.actions, policy.actions)
self.assertEqual(0.5, policy.gamma)
