sess = tf.Session()
actor = Actor(sess, n_features=n_features, n_actions=n_actions, lr=lr_actor)
critic = Critic(sess, n_features=n_features, lr=lr_critic)
sess.run(tf.global_variables_initializer())
for i_episode in range(MAX_EPISODE):
_, state = env.reset()
step = 0
track_r = []
while True:
action = actor.choose_action(state)
_, next_state, reward, done = env.step(action)
env.render()
track_r.append(reward)
td_error = critic.learn(state, reward, next_state)
actor.learn(state, action, td_error)
state = next_state
step += 1
if done or step >= MAX_EP_STEPS:
ep_rs_sum = sum(track_r)
if 'running_reward' not in globals():
running_reward = ep_rs_sum
else:
running_reward = running_reward * 0.95 + ep_rs_sum * 0.05
print("episode:", i_episode, "step:", step, " reward:", int(running_reward))
break
grid_world.create_grid_ui(grid_world.m, grid_world.n,
(grid_world.start_x, grid_world.start_y),
(grid_world.end_x, grid_world.end_y),
grid_world.obstacles)
agent = SARSAgent(actions=list(range(grid_world.action_size)))
number_of_episodes = 10
for episode in range(number_of_episodes):
# reset environment and initialize state
state = grid_world.reset()
# get action of state from agent
action = agent.get_action(str(state))
while True:
grid_world.render()
# take action and proceed one step in the environment
next_state, reward, done = grid_world.step(action)
next_action = agent.get_action(str(next_state))
# with sample <s,a,r,s',a'>, agent learns new q function
agent.learn(str(state), action, reward, str(next_state),
next_action)
state = next_state
action = next_action
# print q function of all states at screen
#env.print_value_all(agent.q_table)
