def td_fa_test():
env = BlackjackEnv()
estimator = Estimator()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
td_fa(env, sess, estimator)
sess.close()
def q_network_test():
env = BlackjackEnv()
estimator = Estimator(0.001)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
V = q_network(env, sess, estimator, episode_num=10000)
plotting.plot_value_function(V, title='Optimal Value Function')
def mc_control_with_epsilon_greedy_test():
env = BlackjackEnv()
Q = mc_control_with_epsilon_greedy(env, episode_nums=10000)
V = defaultdict(float)
for state, actions in Q.items():
max_q = np.max(actions)
V[state] = max_q
plotting.plot_value_function(V, title='Optimal Value Function')
def q_learning_test():
env = BlackjackEnv()
Q = q_learning(env, episode_nums=10000)
V = defaultdict(float)
for state, actions in Q.items():
max_q = np.max(actions)
V[state] = max_q
plotting.plot_value_function(V, title='Optimal Value Function')
def main():
env = BlackjackEnv()
actor = Actor()
estimator = Estimator()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
V = ac_test4debug(sess, env, actor, estimator, episode_num=10000)
plotting.plot_value_function(V, title='Optimal Value Function')
def td_network_test():
env = BlackjackEnv()
estimator = Estimator(learning_rate=0.003)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
V = td_network(env, sess, estimator)
#print(sess.run(estimator.w))
#print(sess.run(estimator.b))
plotting.plot_value_function(V, title='Optimal Value')
def dyna_q_test():
env = BlackjackEnv()
estimator = Estimator(0.003)
model = Model(0.003)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
V = dyna_q(env,
sess,
estimator,
model,
episode_num=3000,
train_model_times=3000,
train_with_model_times=3)
plotting.plot_value_function(V, title='Optimal Value Function')