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 test_feature_gridworld_maxent_irl():
np.random.seed(0)
# env
N = 15
init_pos = np.zeros((N, N), dtype=float)
for i, j in product(range(N // 2 + 2, N), range(N // 2 + 2, N)):
init_pos[i, j] = i**2 + j**2
init_pos /= np.sum(init_pos)
goal_pos = np.array([[n, 0] for n in range(N)])
grid = np.zeros((N, N), dtype=float)
grid = construct_goal_reward(grid, goal_pos, 10)
grid = construct_feature_boundary_reward(
grid,
boundary_axis=0,
boundary_value=N // 2,
reward=-10,
exp_constant=0.2,
)
plot_grid_map(init_pos.T,
"Initial Position Distribution",
cmap=plt.cm.Blues)
plot_grid_map(grid.T, "Reward (Ground Truth)", cmap=plt.cm.Reds)
env = GridWorld(
dimensions=(N, N),
init_pos=init_pos,
goal_pos=goal_pos,
reward_grid=grid,
action_success_rate=1,
render=False,
)
# learn a policy
mdp_algo = value_iteration(env.transition, env.reward, gamma=0.99)
policy = StochasticGreedyPolicy(env.action_space(), mdp_algo,
env.transition)
# roll out trajectories
dataset = collect_trajectories(policy=policy,
env=env,
num_trajectories=20,
maxlen=N * 2)
plot_dataset_distribution(dataset, (N, N), "Dataset State Distribution")
# IRL feature map
feature_map = [
env._feature_map(s) for s in range(env.observation_space().n)
]
feature_map = np.array(feature_map)
# IRL
me_irl = MaxEntIRL(observation_space=env.observation_space(),
action_space=env.action_space(),
transition=env.transition,
goal_states=env.goal_states,
dataset=dataset,
feature_map=feature_map,
max_iter=10,
lr=0.1,
anneal_rate=0.9)
Rprime = me_irl.train()
Rprime = Rprime.reshape((N, N)).T
# plot results
plot_grid_map(Rprime, "Reward (IRL)", cmap=plt.cm.Blues)
plt.show()
def test_gridworld_maxent_irl():
np.random.seed(0)
# env
N = 10
goal_pos = np.array([[N - 1, N - 1]])
human_pos = np.array([[3, 3]])
human_radius = 2
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=False,
)
# learn a policy
mdp_algo = value_iteration(env.transition, env.reward, gamma=0.99)
# mdp_algo = q_learning(env.transition, env.reward, gamma=0.99)
# policy = GreedyPolicy(env.action_space(), mdp_algo)
# policy = EpsGreedyPolicy(env.action_space(), mdp_algo, epsilon=0.1)
policy = StochasticGreedyPolicy(env.action_space(), mdp_algo,
env.transition)
V = np.asarray(mdp_algo.V).reshape((N, N)).T
R = env.reward.reshape((N, N)).T
plot_grid_map(R, "Reward (Ground Truth)", cmap=plt.cm.Reds)
plot_grid_map(V, "Value Function", cmap=plt.cm.Blues)
# roll out trajectories
dataset = collect_trajectories(policy=policy,
env=env,
num_trajectories=200,
maxlen=N * 2)
plot_dataset_distribution(dataset, (N, N), "Dataset State Distribution")
# feature map
feature_map = [
env._feature_map(s) for s in range(env.observation_space().n)
]
feature_map = np.array(feature_map)
# IRL
me_irl = MaxEntIRL(observation_space=env.observation_space(),
action_space=env.action_space(),
transition=env.transition,
goal_states=env.goal_states,
dataset=dataset,
feature_map=feature_map,
max_iter=10,
lr=0.1,
anneal_rate=0.9)
Rprime = me_irl.train()
Rprime = Rprime.reshape((N, N)).T
# plot results
plot_grid_map(Rprime, "Reward (IRL)", print_values=True, cmap=plt.cm.Blues)
plt.show()