def plot(actor):
paths = []
actions_plot = []
env = MultiGoalEnv()
n_games = 50
max_episode_length = 30
for i in range(n_games):
observation = env.reset(init_state=[-3, 0])
episode_length = 0
done = False
score = 0
path = {'infos': {'pos': []}}
while not done:
env.render()
#print('state: ', np.squeeze(observation))
action, _ = actor.forward(T.Tensor([observation]).to(actor.device))
action = action.cpu().detach().numpy()[0]
#print('ac: ', action[0].cpu().detach().numpy())
observation_, reward, done, info = env.step(action)
path['infos']['pos'].append(observation)
if episode_length == max_episode_length:
done = True
episode_length += 1
#print('re:', reward)
score += reward
observation = observation_
paths.append(path)
env.render_rollouts(paths, fout="test_%d.png" % i)
def plot(agent):
paths = []
actions_plot = []
env = MultiGoalEnv()
n_games = 50
max_episode_length = 20
for i in range(n_games):
observation = env.reset(init_state=[0, 0])
episode_length = 0
done = False
score = 0
path = {'infos': {'pos': []}}
while not done:
env.render()
#print('state: ', np.squeeze(observation))
action = agent.choose_action(observation)
#print('ac: ', action[0].cpu().detach().numpy())
observation_, reward, done, info = env.step(action)
path['infos']['pos'].append(observation)
if episode_length == max_episode_length:
done = True
episode_length += 1
#print('re:', reward)
score += reward
observation = observation_
paths.append(path)
score = score / 200
score_history.append(score)
avg_score = np.mean(score_history[-20:])
env.render_rollouts(paths, fout="test_%d.png" % i)
def plot_paths(self, epoch):
paths = []
actions_plot = []
env = MultiGoalEnv()
for episode in range(50):
observation = env.reset()
done = False
step = 0
path = {'infos': {'pos': []}}
particles = None
while not done and step < 30:
self.SVGD_Network.eval()
actions = self.get_sample(observation, 1)
self.SVGD_Network.train()
observation, reward, done, _ = env.step(actions)
path['infos']['pos'].append(observation)
step += 1
paths.append(path)
print("saving figure..., epoch=", epoch)
env.render_rollouts(paths, fout="test_%d.png" % epoch)
#print(env.action_space.high)
best_score = env.reward_range[0]
score_history = []
load_checkpoint = False
max_episode_length = 30
if load_checkpoint:
agent.load_models()
env.render(mode='human')
for i in range(n_games):
observation = env.reset(init_state=[0, 0])
episode_length = 0
done = False
score = 0
while not done:
env.render()
#print('state: ', np.squeeze(observation))
action = agent.choose_action(observation)
#print('ac: ', np.squeeze(action))
observation_, reward, done, info = env.step(action)
if episode_length == max_episode_length:
done = True
episode_length += 1
replay_size=int(1e6),
pi_lr=1e-3,
q_lr=1e-3,
batch_size=100,
n_particles=16,
gamma=0.99,
polyak=0.995)
epochs = 100
update_after = 0
max_ep_len = 30
steps_per_epoch = 400
# Prepare for interaction with environment
total_steps = steps_per_epoch * epochs
o, ep_ret, ep_len = env.reset(), 0, 0
# Main loop: collect experience in env and update/log each epoch
n_particles = 16
epsilon = 0.8
for t in range(total_steps):
if epsilon > 0.2:
epsilon -= 0.00001
#a = agent.get_sample(o)
if np.random.uniform(0, 1) > epsilon:
a = agent.get_sample(o, n_sample=n_particles)
# ind = np.random.choice(np.array([i for i in range(0, n_particles)]))
# a = a[ind]
Q_values = agent.Q_Network(T.tensor(o).float().unsqueeze(0).to(