def test_gridworld_q_learning():
np.random.seed(0)
N = 5
goal_pos = np.array([[N-1, N-1]])
human_pos = np.array([[N-1, 0]])
human_radius = 2
grid = np.ones((N, N), dtype=float) * -1
grid = construct_goal_reward(grid, goal_pos, 10)
grid = construct_human_radius_reward(grid, human_pos, human_radius, -10)
env = GridWorld(
dimensions=(N, N),
init_pos=(0, 0),
goal_pos=goal_pos,
reward_grid=grid,
human_pos=human_pos,
action_success_rate=0.8,
render=True,
)
mdp_algo = q_learning(env.transition, env.reward, gamma=0.99)
mdp_algo.run()
policy = StochasticGreedyPolicy(
env.action_space(), mdp_algo, env.transition)
# plot results
R = env.reward.reshape((N, N)).T
V = np.asarray(mdp_algo.V).reshape((N, N)).T
plot_grid_map(R, "Reward", cmap=plt.cm.Reds)
plot_grid_map(V, "Value Function", cmap=plt.cm.Blues)
plt.show()
obs, rew, done, info = env.reset()
while not done:
act = policy.get_action(obs)
obs, rew, done, info = env.step(act)
time.sleep(0.2)
env.close()
def test_gridworld_value_iteration():
np.random.seed(0)
N = 10
goal_pos = np.array([[N-1, N-1], [N-1, N-2]])
human_pos = np.array([[N//2, N//2], [N-1, 0]])
human_radius = 3
grid = np.zeros((N, N), dtype=float)
grid = construct_goal_reward(grid, goal_pos, 10)
grid = construct_human_radius_reward(grid, human_pos, human_radius, -10)
env = GridWorld(
dimensions=(N, N),
init_pos=(0, 0),
goal_pos=goal_pos,
reward_grid=grid,
human_pos=human_pos,
action_success_rate=1,
render=True,
)
mdp_algo = value_iteration(env.transition, env.reward, gamma=0.99)
policy = EpsGreedyPolicy(env.action_space(), mdp_algo)
# plot results
R = env.reward.reshape((N, N)).T
V = np.asarray(mdp_algo.V).reshape((N, N)).T
plot_grid_map(R, "Reward", cmap=plt.cm.Reds)
plot_grid_map(V, "Value Function", cmap=plt.cm.Blues)
plot_policy(policy, (N, N), "Policy", values=V, cmap=plt.cm.Blues)
plt.show()
obs, rew, done, info = env.reset()
while not done:
act = policy.get_action(obs)
obs, rew, done, info = env.step(act)
time.sleep(0.2)
env.close()
def main(cfg):
pygame.init()
# フォントの作成
sysfont = pygame.font.SysFont(None, 40)
screen = pygame.display.set_mode(WINDOW_SIZE)
pygame.display.set_caption("Grid World")
done = False
clock = pygame.time.Clock()
# grid worldの初期化
grid_env = GridWorld() # grid worldの環境の初期化
ini_state = grid_env.start_pos # 初期状態（エージェントのスタート地点の位置）
agent = QLearningAgent(
epsilon=cfg["agent"]["epsilon"],
epsilon_decay_rate=cfg["agent"]["epsilon_decay_rate"],
actions=np.arange(4),
observation=ini_state) # Q学習エージェント
nb_episode = cfg["nb_episode"] # エピソード数
save_interval = cfg["save_interval"]
result_dir = cfg["result_dir"]
max_step = 1
rewards = [] # 評価用報酬の保存
is_end_episode = False # エージェントがゴールしてるかどうか？
step = 0
# time.sleep(30)
for episode in range(nb_episode):
print("episode:", episode)
episode_reward = [] # 1エピソードの累積報酬
step = 0
while (is_end_episode is False and step < max_step): # ゴールするまで続ける
action = agent.act() # 行動選択
state, reward, is_end_episode = grid_env.step(action)
agent.observe(state, reward) # 状態と報酬の観測
episode_reward.append(reward)
screen.fill(BLACK)
# grid worldの描画
draw_grid_world(grid_env.map, screen)
# テキストを描画したSurfaceを作成
step_str = sysfont.render("step:{}".format(step), False, WHITE)
# 位# テキストを描画する
screen.blit(step_str, (500, 50))
clock.tick(1)
step += 1
# 再描画
pygame.display.flip()
rewards.append(np.sum(episode_reward)) # このエピソードの平均報酬を与える
state = grid_env.reset() # 初期化
agent.observe(state) # エージェントを初期位置に
is_end_episode = False
print("step:", step)
agents = [agent]
if episode % save_interval == 0:
save_result(agents, episode, result_dir)
pygame.quit()
learning_rate=0.0025,
name='dqn_' + str(i),
gamma=0.9,
buffer_size=500, batch_size=10, target_change_step=10, min_buffer_size=100,
load_path=None,
save_path=None
)
for episode in range(EPISODES):
observation = env.state
observation_oh = env.one_hot(env.state)
episode_reward = 0
while True:
action = Agent.act(observation_oh, env.A[observation])
state, reward, done = env.step(action)
state_oh = env.one_hot(state)
Agent.record(observation_oh, action, reward, state_oh, done)
observation = state
observation_oh = state_oh
episode_reward += reward
if done:
reward_list.append(episode_reward)
env.reset()
Agent.learn()
if episode % 100 == 0:
print('Reward for agent %d, episode %d, is %f' % (i, episode, np.mean(reward_list[-10::])))
f_list = []
for test_state in range(env.state_space):
if test_state in env.w:
from grid_world import GridWorld
gw = GridWorld(num_rows=4, num_columns=4)
gw.current_state = 14
print(gw.step('RIGHT'))
for i in range(num_episodes + 1):
episode_buffer = ExperienceBuffer()
observation = env.reset()
observation = np.reshape(observation, [21168])
done = False
total_reward_in_episode = 0
steps_in_episode = 0
while steps_in_episode < max_episode_length:
# 积累样本
steps_in_episode += 1
steps += 1
if np.random.rand(1) < random_threshold or steps < pre_train_steps:
action = np.random.randint(0, 4)
else:
action = sess.run(main_DQN.action, feed_dict={main_DQN.scalar_input: [observation]})[0]
new_observation, reward, done = env.step(action)
new_observation = np.reshape(new_observation, [21168])
#Save the experience to episode buffer
episode_buffer.add([[observation, action, reward, new_observation, done]])
total_reward_in_episode += reward
observation = new_observation
# 积累了 pre_train_steps/max_episode_length 次经历的pre_train_steps个样本之后才会第一次开始衰减随机门限，进行第一次训练。
if steps >= pre_train_steps:
if random_threshold > random_lower_bound:
random_threshold -= drop_step
if steps % (update_freq) == 0:
train_batch = global_experience_buffer.sample(batch_size)
#Below we perform the Double-DQN update to the target Q-values
observations = np.vstack([record[0] for record in train_batch])
actions = np.array([record[1] for record in train_batch])
from policy import EpsGreedyQPolicy
from grid_world import GridWorld
if __name__ == '__main__':
grid_env = GridWorld() # grid worldの環境の初期化
ini_state = grid_env.start_pos # 初期状態（エージェントのスタート地点の位置）
policy = EpsGreedyQPolicy(epsilon=.01) # 方策の初期化。ここではε-greedy
agent = QLearningAgent(actions=np.arange(4), observation=ini_state, policy=policy) # Q Learning エージェントの初期化
nb_episode = 100 #エピソード数
rewards = [] # 評価用報酬の保存
is_goal = False # エージェントがゴールしてるかどうか？
for episode in range(nb_episode):
episode_reward = [] # 1エピソードの累積報酬
while(is_goal == False): # ゴールするまで続ける
action = agent.act() # 行動選択
state, reward, is_goal = grid_env.step(action)
agent.observe(state, reward) # 状態と報酬の観測
episode_reward.append(reward)
rewards.append(np.sum(episode_reward)) # このエピソードの平均報酬を与える
state = grid_env.reset() # 初期化
agent.observe(state) # エージェントを初期位置に
is_goal = False
# テスト(greedyアクション)
agent.traning = False
while(is_goal == False): # ゴールするまで続ける
print("(y, x):{}".format(state))
action = agent.act() # 行動選択
print(action)
state, reward, is_goal = grid_env.step(action)
agent.observe(state, reward) # 状態と報酬の観測
from qlearning_agent import QLearningAgent
from grid_world import GridWorld
if __name__ == '__main__':
grid_env = GridWorld() # grid worldの環境の初期化
ini_state = grid_env.start_pos # 初期状態（エージェントのスタート地点の位置）
agent = QLearningAgent(epsilon=.1,
actions=np.arange(4),
observation=ini_state) # Q学習エージェント
nb_episode = 1000 #エピソード数
rewards = [] # 評価用報酬の保存
is_end_episode = False # エージェントがゴールしてるかどうか？
for episode in range(nb_episode):
episode_reward = [] # 1エピソードの累積報酬
while (is_end_episode == False): # ゴールするまで続ける
action = agent.act() # 行動選択
state, reward, is_end_episode = grid_env.step(action)
agent.observe(state, reward) # 状態と報酬の観測
episode_reward.append(reward)
rewards.append(np.sum(episode_reward)) # このエピソードの平均報酬を与える
state = grid_env.reset() # 初期化
agent.observe(state) # エージェントを初期位置に
is_end_episode = False
# 結果のプロット
plt.plot(np.arange(nb_episode), rewards)
plt.xlabel("episode")
plt.ylabel("reward")
plt.savefig("result.jpg")
plt.show()