def test_agent(env_name, policy_update_filename, frame_stack):
policy_update = torch.load(policy_update_filename)
policy_update.eval()
logger.debug(f"Loaded : {policy_update}")
env = make_env(env_name, 1, frame_stack=frame_stack)
obs = env.reset()
done = np.array([False])
policy = policy_update.policy
total_reward = 0
steps = 0
while not done.all():
env.render()
dist = policy(torch.from_numpy(obs).float())
act = dist.sample()
obs, rew, done, _ = env.step(act.numpy())
total_reward += rew
steps += 1
logger.debug(f"Total reward : {total_reward}")
logger.debug(f"Episode length : {steps}")
env.close()
parser.add_argument("--env-name",
"--env",
type=str,
default="CarRacing-v0")
parser.add_argument("--num-envs",
type=int,
default=multiprocessing.cpu_count())
parser.add_argument("--n-episodes", type=int, default=4)
parser.add_argument("--alpha", type=float, default=0.5)
parser.add_argument("--epochs", type=int, default=1000)
args = parser.parse_args()
logger.info("Using RCRC formulation.")
env = make_env(args.env_name, args.num_envs)
if isinstance(env.action_space, Discrete):
n_acts = env.action_space.n
elif isinstance(env.action_space, Box):
assert (
len(env.action_space.shape) == 1
), f"This example only works for envs with Box(n,) not {env.action_space} action spaces."
n_acts = env.action_space.shape[0]
fixed_model = FixedRandomModel(args.alpha)
update = RCRCUpdate(fixed_model, args.num_envs, n_acts)
update.train()
solve(
args.env_name,
def test_cartpole_v0(self):
env_name = "CartPole-v0"
env = make_env(env_name, 1)
policy_update = self.get_update_model(env)
result = solve(env_name, env, policy_update, logdir)
self.assertEqual(result, True)
def test_lunar_lander_v2(self):
env_name = "LunarLander-v2"
env = make_env(env_name, 1)
policy_update = self.get_update_model(env)
result = solve(env_name, env, policy_update, logdir, epochs=500)
self.assertEqual(result, True)