Agents

Coach supports many state-of-the-art reinforcement learning algorithms, which are separated into three main classes - value optimization, policy optimization and imitation learning. A detailed description of those algorithms can be found by navigating to each of the algorithm pages.

Agents

• Actor-Critic
• ACER
• Behavioral Cloning
• Bootstrapped DQN
• Categorical DQN
• Conditional Imitation Learning
• Clipped Proximal Policy Optimization
• Deep Deterministic Policy Gradient
• Direct Future Prediction
• Double DQN
• Deep Q Networks
• Dueling DQN
• Mixed Monte Carlo
• N-Step Q Learning
• Normalized Advantage Functions
• Neural Episodic Control
• Persistent Advantage Learning
• Policy Gradient
• Proximal Policy Optimization
• Rainbow
• Quantile Regression DQN
• Soft Actor-Critic
• Twin Delayed Deep Deterministic Policy Gradient
• Wolpertinger

class rl_coach.base_parameters.AgentParameters(algorithm: rl_coach.base_parameters.AlgorithmParameters, exploration: ExplorationParameters, memory: MemoryParameters, networks: Dict[str, rl_coach.base_parameters.NetworkParameters], visualization: rl_coach.base_parameters.VisualizationParameters = <rl_coach.base_parameters.VisualizationParameters object>)

Parameters

• algorithm – A class inheriting AlgorithmParameters. The parameters used for the specific algorithm used by the agent. These parameters can be later referenced in the agent implementation through self.ap.algorithm.

• exploration – Either a class inheriting ExplorationParameters or a dictionary mapping between action space types and their corresponding ExplorationParameters. If a dictionary was used, when the agent will be instantiated, the correct exploration policy parameters will be used according to the real type of the environment action space. These parameters will be used to instantiate the exporation policy.

• memory – A class inheriting MemoryParameters. It defines all the parameters used by the memory module.

• networks – A dictionary mapping between network names and their corresponding network parmeters, defined as a class inheriting NetworkParameters. Each element will be used in order to instantiate a NetworkWrapper class, and all the network wrappers will be stored in the agent under self.network_wrappers. self.network_wrappers is a dict mapping between the network name that was given in the networks dict, and the instantiated network wrapper.

• visualization – A class inheriting VisualizationParameters and defining various parameters that can be used for visualization purposes, such as printing to the screen, rendering, and saving videos.

class rl_coach.agents.agent.Agent(agent_parameters: rl_coach.base_parameters.AgentParameters, parent: Union[LevelManager, CompositeAgent] = None)

Parameters

agent_parameters – A AgentParameters class instance with all the agent parameters

act(action: Union[None, int, float, numpy.ndarray, List] = None) → rl_coach.core_types.ActionInfo

Given the agents current knowledge, decide on the next action to apply to the environment

Parameters

action – An action to take, overriding whatever the current policy is

Returns

An ActionInfo object, which contains the action and any additional info from the action decision process

call_memory(func, args=())

This function is a wrapper to allow having the same calls for shared or unshared memories. It should be used instead of calling the memory directly in order to allow different algorithms to work both with a shared and a local memory.

Parameters

• func – the name of the memory function to call

• args – the arguments to supply to the function

Returns

the return value of the function

choose_action(curr_state)

choose an action to act with in the current episode being played. Different behavior might be exhibited when training or testing.

Parameters

curr_state – the current state to act upon.

Returns

chosen action, some action value describing the action (q-value, probability, etc)

collect_savers(parent_path_suffix: str) → rl_coach.saver.SaverCollection

Collect all of agent’s network savers :param parent_path_suffix: path suffix of the parent of the agent (could be name of level manager or composite agent) :return: collection of all agent savers

create_networks() → Dict[str, rl_coach.architectures.network_wrapper.NetworkWrapper]

Create all the networks of the agent. The network creation will be done after setting the environment parameters for the agent, since they are needed for creating the network.

Returns

A list containing all the networks

freeze_memory()

Shuffle episodes in the memory and freeze it to make sure that no extra data is being pushed anymore. :return: None

get_predictions(states: List[Dict[str, numpy.ndarray]], prediction_type: rl_coach.core_types.PredictionType)

Get a prediction from the agent with regard to the requested prediction_type. If the agent cannot predict this type of prediction_type, or if there is more than possible way to do so, raise a ValueException.

Parameters

• states – The states to get a prediction for

• prediction_type – The type of prediction to get for the states. For example, the state-value prediction.

Returns

the predicted values

get_state_embedding(state: dict) → numpy.ndarray

Given a state, get the corresponding state embedding from the main network

Parameters

state – a state dict

Returns

a numpy embedding vector

handle_episode_ended() → None

Make any changes needed when each episode is ended. This includes incrementing counters, updating full episode dependent values, updating logs, etc. This function is called right after each episode is ended.

Returns

None

init_environment_dependent_modules() → None

Initialize any modules that depend on knowing information about the environment such as the action space or the observation space

Returns

None

initialize_session_dependent_components()

Initialize components which require a session as part of their initialization.

Returns

None

learn_from_batch(batch) → Tuple[float, List, List]

Given a batch of transitions, calculates their target values and updates the network.

Parameters

batch – A list of transitions

Returns

The total loss of the training, the loss per head and the unclipped gradients

load_memory_from_file()

Load memory transitions from a file.

Returns

None

log_to_screen() → None

Write an episode summary line to the terminal

Returns

None

observe(env_response: rl_coach.core_types.EnvResponse) → bool

Given a response from the environment, distill the observation from it and store it for later use. The response should be a dictionary containing the performed action, the new observation and measurements, the reward, a game over flag and any additional information necessary.

Parameters

env_response – result of call from environment.step(action)

Returns

a boolean value which determines if the agent has decided to terminate the episode after seeing the given observation

property parent

Get the parent class of the agent

Returns

the current phase

property phase

The current running phase of the agent

Returns

RunPhase

post_training_commands() → None

A function which allows adding any functionality that is required to run right after the training phase ends.

Returns

None

prepare_batch_for_inference(states: Union[Dict[str, numpy.ndarray], List[Dict[str, numpy.ndarray]]], network_name: str) → Dict[str, numpy.core.multiarray.array]

Convert curr_state into input tensors tensorflow is expecting. i.e. if we have several inputs states, stack all observations together, measurements together, etc.

Parameters

• states – A list of environment states, where each one is a dict mapping from an observation name to its corresponding observation

• network_name – The agent network name to prepare the batch for. this is needed in order to extract only the observation relevant for the network from the states.

Returns

A dictionary containing a list of values from all the given states for each of the observations

register_signal(signal_name: str, dump_one_value_per_episode: bool = True, dump_one_value_per_step: bool = False) → rl_coach.utils.Signal

Register a signal such that its statistics will be dumped and be viewable through dashboard

Parameters

• signal_name – the name of the signal as it will appear in dashboard

• dump_one_value_per_episode – should the signal value be written for each episode?

• dump_one_value_per_step – should the signal value be written for each step?

Returns

the created signal

reset_evaluation_state(val: rl_coach.core_types.RunPhase) → None

Perform accumulators initialization when entering an evaluation phase, and signal dumping when exiting an evaluation phase. Entering or exiting the evaluation phase is determined according to the new phase given by val, and by the current phase set in self.phase.

Parameters

val – The new phase to change to

Returns

None

reset_internal_state() → None

Reset all the episodic parameters. This function is called right before each episode starts.

Returns

None

restore_checkpoint(checkpoint_dir: str) → None

Allows agents to store additional information when saving checkpoints.

Parameters

checkpoint_dir – The checkpoint dir to restore from

Returns

None

run_off_policy_evaluation() → None

Run off-policy evaluation estimators to evaluate the trained policy performance against a dataset. Should only be implemented for off-policy RL algorithms.

Returns

None

run_pre_network_filter_for_inference(state: Dict[str, numpy.ndarray], update_filter_internal_state: bool = True) → Dict[str, numpy.ndarray]

Run filters which where defined for being applied right before using the state for inference.

Parameters

• state – The state to run the filters on

• update_filter_internal_state – Should update the filter’s internal state - should not update when evaluating

Returns

The filtered state

save_checkpoint(checkpoint_prefix: str) → None

Allows agents to store additional information when saving checkpoints.

Parameters

checkpoint_prefix – The prefix of the checkpoint file to save

Returns

None

set_environment_parameters(spaces: rl_coach.spaces.SpacesDefinition)

Sets the parameters that are environment dependent. As a side effect, initializes all the components that are dependent on those values, by calling init_environment_dependent_modules

Parameters

spaces – the environment spaces definition

Returns

None

set_incoming_directive(action: Union[int, float, numpy.ndarray, List]) → None

Allows setting a directive for the agent to follow. This is useful in hierarchy structures, where the agent has another master agent that is controlling it. In such cases, the master agent can define the goals for the slave agent, define its observation, possible actions, etc. The directive type is defined by the agent in-action-space.

Parameters

action – The action that should be set as the directive

Returns

set_session(sess) → None

Set the deep learning framework session for all the agents in the composite agent

Returns

None

setup_logger() → None

Setup the logger for the agent

Returns

None

sync() → None

Sync the global network parameters to local networks

Returns

None

train() → float

Check if a training phase should be done as configured by num_consecutive_playing_steps. If it should, then do several training steps as configured by num_consecutive_training_steps. A single training iteration: Sample a batch, train on it and update target networks.

Returns

The total training loss during the training iterations.

update_log() → None

Updates the episodic log file with all the signal values from the most recent episode. Additional signals for logging can be set by the creating a new signal using self.register_signal, and then updating it with some internal agent values.

Returns

None

update_step_in_episode_log() → None

Updates the in-episode log file with all the signal values from the most recent step.

Returns

None

update_transition_before_adding_to_replay_buffer(transition: rl_coach.core_types.Transition) → rl_coach.core_types.Transition

Allows agents to update the transition just before adding it to the replay buffer. Can be useful for agents that want to tweak the reward, termination signal, etc.

Parameters

transition – the transition to update

Returns

the updated transition