def test_ppo_agent_play(env_name):
"""
Extension of the agent set up and initialisation test to include playing episodes.
"""
# Set up environment using default parameters.
# Environment parameters do not affect the test result here.
tf_env, action_dims = rl_env_from_snc_env(load_scenario(
env_name, job_gen_seed=10)[1],
discount_factor=0.99,
normalise_observations=False)
# Instantiate and initialise a PPO agent for the environment.
ppo_agent = create_ppo_agent(tf_env, num_epochs=10)
ppo_agent.initialize()
# Reset the environment
tf_env.reset()
# Play 5 time steps in the environment.
for _ in range(5):
# Since we do not have the state stored at this point we capture it from the environment
# fresh each time step as a TimeStep object (a named tuple).
time_step = tf_env.current_time_step()
# Attain our agent's action.
action_step = ppo_agent.collect_policy.action(time_step)
# Ensure that the action is one-hot as expected
if isinstance(action_step.action, tuple):
action = tf.concat(action_step.action, axis=-1)
else:
action = action_step.action
# Ensure that the action is binary as expected.
assert snc.is_binary(action)
# Play the action out in the environment.
tf_env.step(action_step.action)
def test_ppo_agent_init_with_multiple_resource_sets():
"""
Tests agent set up and initialisation with multiple action subspaces (multiple resource sets).
"""
# Set the environment name for this case as the asserts are difficult to make as variables.
env_name = 'double_reentrant_line_shared_res_homogeneous_cost'
# Set up the environment parameters.
# Environment parameters do not affect the test result here.
tf_env, _ = rl_env_from_snc_env(load_scenario(env_name,
job_gen_seed=10)[1],
discount_factor=0.99,
normalise_observations=False)
# Instantiate and initialise a PPO agent for the environment.
ppo_agent = create_ppo_agent(tf_env, num_epochs=10)
ppo_agent.initialize()
# Validate initialisation by checking some properties of the initalised agent.
assert isinstance(ppo_agent.action_spec, tuple)
assert len(ppo_agent.action_spec) == 2
assert isinstance(ppo_agent.action_spec[0], BoundedTensorSpec)
assert isinstance(ppo_agent.action_spec[1], BoundedTensorSpec)
assert ppo_agent.action_spec[0].shape == tf.TensorShape((1, 3))
assert ppo_agent.action_spec[1].shape == tf.TensorShape((1, 3))
assert ppo_agent.name == "PPO_Agent"
assert ppo_agent.time_step_spec == tf_env.time_step_spec()
def get_ppo_agent(
env: TFPyEnvironment,
num_epochs: int,
discount_factor: float,
debug: bool = False,
agent_params: Optional[Dict[str, Any]] = None
) -> Union[ReinforceAgent, PPOAgent]:
"""
Builds and initialises a reinforcement learning agent for the environment.
:param env: The TensorFlow environment used to set up the agent with correct action spaces etc.
:param num_epochs: The (maximal) number of internal PPO epochs to run.
:param discount_factor: The discount applied to future rewards.
:param debug: Flag which determines whether to include extra TensorBoard logs for debugging.
:param agent_params: A dictionary of possible overrides for the default TF-Agents agent set up.
:return: An initialised RL agent.
"""
# Set up a training step counter.
global_step = tf.compat.v1.train.get_or_create_global_step()
agent = create_ppo_agent(
env,
num_epochs,
gamma=discount_factor,
debug=debug,
training_step_counter=global_step,
agent_params=agent_params
)
agent.initialize()
agent.train = tf.function(agent.train)
return agent
def test_rl_simulation_agent_normalise_obs_property():
"""Ensure that the _normalise_obs property of RLSimulationAgent is set correctly."""
# Set up the agent as before.
seed = 72
env = load_scenario("single_server_queue", job_gen_seed=seed).env
rl_env, _ = rl_env_from_snc_env(env,
discount_factor=0.99,
normalise_observations=False)
ppo_agent = create_ppo_agent(rl_env, gamma=0.90)
ppo_agent.initialize()
del rl_env
ppo_sim_agent = RLSimulationAgent(env, ppo_agent, normalise_obs=False)
assert ppo_sim_agent._normalise_obs is False
ppo_sim_agent = RLSimulationAgent(env, ppo_agent, normalise_obs=True)
assert ppo_sim_agent._normalise_obs is True
def test_rl_simulation_agent_discount_factor_ppo():
"""
Tests that the discount factor is passed from a PPO agent to an RLSimulationAgent correctly.
"""
# Set up the agent as before.
seed = 72
env = load_scenario("single_server_queue", job_gen_seed=seed).env
rl_env, _ = rl_env_from_snc_env(env,
discount_factor=0.99,
normalise_observations=False)
ppo_agent = create_ppo_agent(rl_env, gamma=0.90)
ppo_agent.initialize()
del rl_env
ppo_sim_agent = RLSimulationAgent(env, ppo_agent, normalise_obs=False)
assert ppo_sim_agent.discount_factor == 0.90
def test_ppo_agent_init(env_name, expected_action_spec_shape):
"""
Tests agent set up and initialisation.
"""
# Set up environment using default parameters.
# Environment parameters do not affect the test result here.
tf_env, _ = rl_env_from_snc_env(load_scenario(env_name,
job_gen_seed=10)[1],
discount_factor=0.99,
normalise_observations=False)
# Instantiate and initialise a PPO agent for the environment.
ppo_agent = create_ppo_agent(tf_env, num_epochs=10)
ppo_agent.initialize()
# Validate initialisation by checking relevant properties of the initalised agent.
assert isinstance(ppo_agent.action_spec, BoundedTensorSpec)
assert ppo_agent.action_spec.shape == expected_action_spec_shape
assert ppo_agent.name == "PPO_Agent"
assert ppo_agent.time_step_spec == tf_env.time_step_spec()
def test_ppo_agent_learning(env_name):
"""
Extension of the play test for an agent playing in the environment to include training.
Note: This does not test that training improves the policy. It simply tests that the training
loop runs effectively and changes the policy parameters.
"""
# Set up environment using default parameters.
# Environment parameters do not affect the test result here.
tf_env, _ = rl_env_from_snc_env(load_scenario(
env_name,
job_gen_seed=10,
override_env_params={'max_episode_length': 25})[1],
discount_factor=0.99,
normalise_observations=False)
# Set up a training step counter.
global_step = tf.compat.v1.train.get_or_create_global_step()
# Instantiate a PPO agent
ppo_agent = create_ppo_agent(tf_env,
num_epochs=10,
training_step_counter=global_step)
# Instantiate a replay buffer.
replay_buffer = TFUniformReplayBuffer(
data_spec=ppo_agent.collect_data_spec,
batch_size=tf_env.batch_size,
max_length=1000)
# Use a driver to handle data collection for the agent. This handles a lot of the backend
# TensorFlow set up and solves previous errors with episodes of differing lengths.
collect_driver = DynamicEpisodeDriver(tf_env,
ppo_agent.collect_policy,
observers=[replay_buffer.add_batch],
num_episodes=2)
# collect_driver.run = tf.function(collect_driver.run)
# Get the initial states of the agent and environment before training.
time_step = tf_env.reset()
policy_state = ppo_agent.collect_policy.get_initial_state(
tf_env.batch_size)
# Take a copy of the variables in order to ensure that training does lead to parameter changes.
initial_vars = deepcopy(ppo_agent.trainable_variables)
assert len(initial_vars) > 0, "Agent has no trainable variables."
# Set up a minimal training loop to simply test training mechanics work.
for _ in range(5):
# Collect experience.
time_step, policy_state = collect_driver.run(time_step=time_step,
policy_state=policy_state)
# Now the replay buffer should have data in it so we can collect the data and train the
# agent.
experience = replay_buffer.gather_all()
ppo_agent.train(experience)
# Clear the replay buffer and return to play.
replay_buffer.clear()
# Check that training has had some effect
for v1, v2 in zip(initial_vars, ppo_agent.trainable_variables):
assert not np.allclose(v1.numpy(), v2.numpy())
def load_rl_agent(
env: ControlledRandomWalk,
rl_algorithm: str,
load_path: str,
discount_factor: float = 0.99,
agent_params: Optional[Dict[str, Any]] = None) -> RLSimulationAgent:
"""
Instantiates an RL agent in the RLSimulationAgent interface for compatibility and loads the
weights from training into it.
:param env: The controlled random walk environment for which the agent is required.
:param rl_algorithm: The name of the RL algorithm used to train the agent.
:param load_path: Path to a directory where TensorFlow checkpoints have been saved (i.e. where
the model's weights are saved).
:param discount_factor: A scalar discount factor to pass to the agent.
:param agent_params: A dictionary of possible overrides for the default TF-Agents agent set up.
:return: An RL agent initialised with saved weights ready for evaluation.
"""
# Lazy import of TensorFlow as if no RL agent is run then it isn't needed.
import tensorflow as tf
# Attain a TensorFlow compatible version of the environment.
# We need a TensorFlow environment to initialise the agent correctly.
# First determine whether or not to normalise observations, PPO has its own normalisation so we
# only normalise for reinforce agents or PPO agents where normalisation is turned off.
normalise_obs = rl_algorithm == 'reinforce' or \
(rl_algorithm == 'ppo' and not agent_params.get('normalize_observations', True))
tf_env, _ = rl_env.rl_env_from_snc_env(
env, discount_factor, normalise_observations=normalise_obs)
# Set up an enumeration of functions which build agents to allow for extending to new agents.
# Pick out the correct RL agent from those we have implemented.
if rl_algorithm.lower() == 'reinforce':
agent = create_reinforce_agent(tf_env,
gamma=discount_factor,
agent_params=agent_params)
elif rl_algorithm.lower() == 'ppo':
agent = create_ppo_agent(tf_env,
gamma=discount_factor,
agent_params=agent_params)
else:
raise NotImplementedError(
"An agent using the RL algorithm requested is not yet implemented")
# Initialise the agent and load in parameters from the most recent save.
# Note that this can be adjusted to load in weights from any point in training (so long as they
# have been saved).
agent.initialize()
restorer = tf.train.Checkpoint(agent=agent)
restore_manager = tf.train.CheckpointManager(restorer,
directory=load_path,
max_to_keep=20)
restorer.listed = agent.trainable_variables
restoration = restorer.restore(restore_manager.latest_checkpoint)
restoration.run_restore_ops()
# Check that the weights have been loaded and that the model from which the weights were saved
# matches the model which they are being loaded into.
restoration.assert_nontrivial_match()
restoration.assert_existing_objects_matched()
# We name the agent in line with the checkpoint used to restore the weights. This aids in
# identifying which experiment run is being looked at from log files.
agent_name = f"RLSimulationAgent - {restore_manager.latest_checkpoint}"
# Finally wrap the agent for compatibility with the SNC simulator.
simulation_agent = RLSimulationAgent(env,
agent,
normalise_obs,
name=agent_name)
return simulation_agent