def test_multi_gpu_ppo_agent_learning_test_gridworld_2x2(self):
"""
Tests if the multi gpu strategy can learn successfully on a multi gpu system, but
also runs on a CPU-only system using fake-GPU logic for testing purposes.
"""
env_spec = dict(type="grid-world", world="2x2")
dummy_env = GridWorld.from_spec(env_spec)
agent_config = config_from_path(
"configs/multi_gpu_ppo_for_2x2_gridworld.json")
preprocessing_spec = agent_config.pop("preprocessing_spec")
agent = PPOAgent.from_spec(
agent_config,
state_space=self.grid_world_2x2_flattened_state_space,
action_space=dummy_env.action_space,
)
time_steps = 10000
worker = SingleThreadedWorker(env_spec=env_spec,
agent=agent,
worker_executes_preprocessing=True,
preprocessing_spec=preprocessing_spec)
results = worker.execute_timesteps(time_steps, use_exploration=True)
# Assume we have learned something.
# TODO: This test needs more tuning. -1.0 is not great for the 2x2 grid world.
self.assertGreater(results["mean_episode_reward"], -1.0)
def test_ppo_on_cart_pole(self):
"""
Creates a PPO Agent and runs it via a Runner on the CartPole Env.
"""
env = OpenAIGymEnv("CartPole-v0", seed=36)
agent = PPOAgent.from_spec(
config_from_path("configs/ppo_agent_for_cartpole.json"),
state_space=env.state_space,
action_space=env.action_space
)
time_steps = 3000
worker = SingleThreadedWorker(
env_spec=lambda: env,
agent=agent,
worker_executes_preprocessing=False,
render=self.is_windows
)
results = worker.execute_timesteps(time_steps, use_exploration=True)
print(results)
self.assertEqual(results["timesteps_executed"], time_steps)
self.assertEqual(results["env_frames"], time_steps)
#self.assertGreaterEqual(results["mean_episode_reward"], 23)
#self.assertGreaterEqual(results["max_episode_reward"], 100.0)
self.assertLessEqual(results["episodes_executed"], time_steps / 10)
def test_ppo_on_2x2_grid_world_with_container_actions(self):
"""
Creates a PPO agent and runs it via a Runner on a simple 2x2 GridWorld using container actions.
"""
# ftj = forward + turn + jump
env_spec = dict(world="2x2",
action_type="ftj",
state_representation="xy+orientation")
dummy_env = GridWorld.from_spec(env_spec)
agent_config = config_from_path(
"configs/ppo_agent_for_2x2_gridworld_with_container_actions.json")
preprocessing_spec = agent_config.pop("preprocessing_spec")
agent = PPOAgent.from_spec(agent_config,
state_space=FloatBox(shape=(4, )),
action_space=dummy_env.action_space)
time_steps = 10000
worker = SingleThreadedWorker(
env_spec=lambda: GridWorld.from_spec(env_spec),
agent=agent,
preprocessing_spec=preprocessing_spec,
worker_executes_preprocessing=True,
render=False)
results = worker.execute_timesteps(time_steps, use_exploration=True)
print("Results =", results)
def test_ppo_on_2x2_grid_world(self):
"""
Creates a PPO Agent and runs it via a Runner on the 2x2 Grid World Env.
"""
env = GridWorld(world="2x2")
agent = PPOAgent.from_spec(
config_from_path("configs/ppo_agent_for_2x2_gridworld.json"),
state_space=GridWorld.grid_world_2x2_flattened_state_space,
action_space=env.action_space,
execution_spec=dict(seed=15),
)
time_steps = 3000
worker = SingleThreadedWorker(
env_spec=lambda: env,
agent=agent,
worker_executes_preprocessing=True,
preprocessing_spec=GridWorld.grid_world_2x2_preprocessing_spec
)
results = worker.execute_timesteps(time_steps, use_exploration=True)
print(results)
# Assume we have learned something.
self.assertGreater(results["mean_episode_reward"], -0.2)
def test_ppo_on_cart_pole(self):
"""
Creates a PPO Agent and runs it via a Runner on the CartPole env.
"""
env = OpenAIGymEnv("CartPole-v0", seed=36)
agent = PPOAgent.from_spec(
config_from_path("configs/ppo_agent_for_cartpole.json"),
state_space=env.state_space,
action_space=env.action_space)
time_steps = 3000
worker = SingleThreadedWorker(
env_spec=lambda: env,
agent=agent,
worker_executes_preprocessing=False,
render=False, #self.is_windows
episode_finish_callback=lambda episode_return, duration, timesteps,
env_num: print("episode return {}; steps={}".format(
episode_return, timesteps)))
results = worker.execute_timesteps(time_steps, use_exploration=True)
print(results)
self.assertEqual(results["timesteps_executed"], time_steps)
self.assertEqual(results["env_frames"], time_steps)
self.assertLessEqual(results["episodes_executed"], time_steps / 10)
# Assume we have learned something.
self.assertGreaterEqual(results["mean_episode_reward"], 40.0)
def test_post_processing(self):
"""
Tests external batch post-processing for the PPO agent.
"""
env = OpenAIGymEnv("Pong-v0",
frameskip=4,
max_num_noops=30,
episodic_life=True)
agent_config = config_from_path("configs/ppo_agent_for_pong.json")
agent = PPOAgent.from_spec(agent_config,
state_space=env.state_space,
action_space=env.action_space)
num_samples = 200
states = agent.preprocessed_state_space.sample(num_samples)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
sequence_indices_space = BoolBox(add_batch_rank=True)
# GAE is separately tested, just testing if this API method returns results.
pg_advantages = agent.post_process(
dict(states=states,
rewards=reward_space.sample(num_samples),
terminals=terminal_space.sample(num_samples, fill_value=0),
sequence_indices=sequence_indices_space.sample(num_samples,
fill_value=0)))
def test_value_function_weights(self):
"""
Tests changing of value function weights.
"""
env = OpenAIGymEnv("Pong-v0")
agent_config = config_from_path("configs/ppo_agent_for_pong.json")
agent = PPOAgent.from_spec(agent_config,
state_space=env.state_space,
action_space=env.action_space)
weights = agent.get_weights()
assert "value_function_weights" in weights
assert "policy_weights" in weights
policy_weights = weights["policy_weights"]
value_function_weights = weights["value_function_weights"]
# Just change vf weights.
for key, weight in value_function_weights.items():
value_function_weights[key] = weight + 0.01
agent.set_weights(policy_weights, value_function_weights)
new_actual_weights = agent.get_weights()
recursive_assert_almost_equal(
new_actual_weights["value_function_weights"],
value_function_weights)
def test_ppo_compilation(self):
"""
Tests PPO agent compilation.
"""
env = OpenAIGymEnv("Pong-v0", frameskip=4, max_num_noops=30, episodic_life=True)
agent_config = config_from_path("configs/ppo_agent_for_pong.json")
agent = PPOAgent.from_spec(
agent_config,
state_space=env.state_space,
action_space=env.action_space
)
print("Compiled {}".format(agent))
def test_ppo_on_pendulum(self):
"""
Creates a PPO Agent and runs it via a Runner on the Pendulum env.
"""
env = OpenAIGymEnv("Pendulum-v0")
agent = PPOAgent.from_spec(
config_from_path("configs/ppo_agent_for_pendulum.json"),
state_space=env.state_space,
action_space=env.action_space)
worker = SingleThreadedWorker(env_spec=lambda: env,
agent=agent,
worker_executes_preprocessing=False,
render=self.is_windows)
results = worker.execute_episodes(500, use_exploration=True)
print(results)
def test_ppo_on_pendulum(self):
"""
Creates a PPO Agent and runs it via a Runner on the Pendulum env.
"""
env = OpenAIGymEnv("Pendulum-v0")
agent = PPOAgent.from_spec(
config_from_path("configs/ppo_agent_for_pendulum.json"),
state_space=env.state_space,
action_space=env.action_space)
worker = SingleThreadedWorker(
env_spec=lambda: env,
agent=agent,
worker_executes_preprocessing=False,
render=False, #self.is_windows,
episode_finish_callback=lambda episode_return, duration, timesteps,
env_num: print("episode return {}; steps={}".format(
episode_return, timesteps)))
results = worker.execute_episodes(5000, use_exploration=True)
print(results)
def test_ppo_on_2x2_grid_world_with_container_actions(self):
"""
Creates a PPO agent and runs it via a Runner on a simple 2x2 GridWorld using container actions.
"""
# -----
# |^|H|
# -----
# | |G| ^=start, looking up
# -----
# ftj = forward + turn + jump
env_spec = dict(world="2x2",
action_type="ftj",
state_representation="xy+orientation")
dummy_env = GridWorld.from_spec(env_spec)
agent_config = config_from_path(
"configs/ppo_agent_for_2x2_gridworld_with_container_actions.json")
preprocessing_spec = agent_config.pop("preprocessing_spec")
agent = PPOAgent.from_spec(agent_config,
state_space=FloatBox(shape=(4, )),
action_space=dummy_env.action_space)
time_steps = 5000
worker = SingleThreadedWorker(
env_spec=lambda: GridWorld.from_spec(env_spec),
agent=agent,
preprocessing_spec=preprocessing_spec,
worker_executes_preprocessing=True,
render=False)
results = worker.execute_timesteps(time_steps, use_exploration=True)
print(results)
self.assertEqual(results["timesteps_executed"], time_steps)
self.assertEqual(results["env_frames"], time_steps)
self.assertLessEqual(results["episodes_executed"], time_steps)
# Assume we have learned something.
self.assertGreaterEqual(results["mean_episode_reward"], -2.0)
def test_ppo_on_container_state_and_action_spaces_and_very_large_rewards(
self):
"""
Tests stability of PPO on an extreme env producing strange container states and large rewards and requiring
container actions.
"""
env = RandomEnv(
state_space=Dict(
{"F_position": FloatBox(shape=(2, ), low=0.01, high=0.02)}),
action_space=Dict({
"F_direction_low-1.0_high1.0":
FloatBox(shape=(), low=-1.0, high=1.0),
"F_forward_direction_low-1.0_high1.0":
FloatBox(shape=(), low=-1.0, high=1.0),
"B_jump":
BoolBox()
}),
reward_space=FloatBox(low=-1000.0,
high=-100000.0), # hugely negative rewards
terminal_prob=0.0000001)
agent_config = config_from_path(
"configs/ppo_agent_for_random_env_with_container_spaces.json")
agent = PPOAgent.from_spec(agent_config,
state_space=env.state_space,
action_space=env.action_space)
worker = SingleThreadedWorker(
env_spec=lambda: env,
agent=agent,
preprocessing_spec=None,
worker_executes_preprocessing=True,
#episode_finish_callback=lambda episode_return, duration, timesteps, env_num:
#print("episode return {}; steps={}".format(episode_return, timesteps))
)
results = worker.execute_timesteps(num_timesteps=int(1e6),
use_exploration=True)
print(results)