Agent and model overview

A reinforcement learning agent provides methods to process states and return actions, to store past observations, and to load and save models. Most agents employ a Model which implements the algorithms to calculate the next action given the current state and to update model parameters from past experiences.

Environment <-> Runner <-> Agent <-> Model

Parameters to the agent are passed as a Configuration object. The configuration is passed on to the Model.

Ready-to-use algorithms

We implemented some of the most common RL algorithms and try to keep these up-to-date. Here we provide an overview over all implemented agents and models.

Agent / General parameters

Agent is the base class for all reinforcement learning agents. Every agent inherits from this class.

class tensorforce.agents.Agent(states_spec, actions_spec, config)

Basic Reinforcement learning agent. An agent encapsulates execution logic of a particular reinforcement learning algorithm and defines the external interface to the environment.

The agent hence acts an intermediate layer between environment and backend execution (value function or policy updates).

Each agent requires the following configuration parameters:

  • states: dict containing one or more state definitions.
  • actions: dict containing one or more action definitions.
  • preprocessing: dict or list containing state preprocessing configuration.
  • exploration: dict containing action exploration configuration.

The configuration is passed to the Model and should thus include its configuration parameters, too.

Examples:

act(states, deterministic=False)

Return action(s) for given state(s). First, the states are preprocessed using the given preprocessing configuration. Then, the states are passed to the model to calculate the desired action(s) to execute.

After obtaining the actions, exploration might be added by the agent, depending on the exploration configuration.

Parameters:
  • states – One state (usually a value tuple) or dict of states if multiple states are expected.
  • deterministic – If true, no exploration and sampling is applied.
Returns:

Scalar value of the action or dict of multiple actions the agent wants to execute.
static from_spec(spec, kwargs)

Creates an agent from a specification dict.

observe(terminal, reward)

Observe experience from the environment to learn from. Optionally preprocesses rewards Child classes should call super to get the processed reward EX: terminal, reward = super()...

Parameters:
  • terminal – boolean indicating if the episode terminated after the observation.
  • reward – scalar reward that resulted from executing the action.
reset()

Reset the agent to its initial state on episode start. Updates internal episode and
 timestep counter, internal states, and resets preprocessors.

restore_model(directory=None, file=None)

Restore TensorFlow model. If no checkpoint file is given, the latest checkpoint is
 restored. If no checkpoint directory is given, the model’s default saver directory is
 used (unless file specifies the entire path).

Parameters:
  • directory – Optional checkpoint directory.
  • file – Optional checkpoint file, or path if directory not given.
save_model(directory=None, append_timestep=True)

Save TensorFlow model. If no checkpoint directory is given, the model’s default saver
 directory is used. Optionally appends current timestep to prevent overwriting previous
 checkpoint files. Turn off to be able to load model from the same given path argument as
 given here.

Parameters:
  • directory – Optional checkpoint directory.
  • use_global_step – Appends the current timestep to the checkpoint file if true.

:param If this is set to True, the load path must include the checkpoint timestep suffix.
: :param For example, if stored to models/ and set to true, the exported file will be of the
: :param form models/model.ckpt-X where X is the last timestep saved. The load path must
: :param precisely match this file name. If this option is turned off, the checkpoint will
: :param always overwrite the file specified in path and the model can always be loaded under
: :param this path.:

Returns:Checkpoint path were the model was saved.

Model

The Model class is the base class for reinforcement learning models.

class tensorforce.models.Model(states_spec, actions_spec, config, **kwargs)

Bases: object

Base class for all (TensorFlow-based) models.

create_output_operations(states, internals, actions, terminal, reward, update, deterministic)

Calls all the relevant TensorFlow functions for this model and hence creates all the TensorFlow operations involved.

Parameters:
  • states – Dict of state tensors.
  • internals – List of prior internal state tensors.
  • actions – Dict of action tensors.
  • terminal – Terminal boolean tensor.
  • reward – Reward tensor.
  • update – Boolean tensor indicating whether this call happens during an update.
  • deterministic – Boolean tensor indicating whether action should be chosen
     deterministically.
get_optimizer_kwargs(states, internals, actions, terminal, reward, update)

Returns the optimizer arguments including the time, the list of variables to optimize, and various argument-free functions (in particular fn_loss returning the combined 0-dim batch loss tensor) which the optimizer might require to perform an update step.

Parameters:
  • states – Dict of state tensors.
  • internals – List of prior internal state tensors.
  • actions – Dict of action tensors.
  • terminal – Terminal boolean tensor.
  • reward – Reward tensor.
  • update – Boolean tensor indicating whether this call happens during an update.
Returns:

Loss tensor of the size of the batch.
get_summaries()

Returns the TensorFlow summaries reported by the model

Returns:List of summaries
get_variables(include_non_trainable=False)

Returns the TensorFlow variables used by the model.

Returns:List of variables.
initialize(custom_getter)

Creates the TensorFlow placeholders and functions for this model. Moreover adds the internal state placeholders and initialization values to the model.

Parameters:custom_getter – The custom_getter_ object to use for tf.make_template when creating TensorFlow functions.
reset()

Resets the model to its initial state on episode start.

Returns:Current episode and timestep counter, and a list containing the internal states
 initializations.
restore(directory=None, file=None)

Restore TensorFlow model. If no checkpoint file is given, the latest checkpoint is
 restored. If no checkpoint directory is given, the model’s default saver directory is
 used (unless file specifies the entire path).

Parameters:
  • directory – Optional checkpoint directory.
  • file – Optional checkpoint file, or path if directory not given.
save(directory=None, append_timestep=True)

Save TensorFlow model. If no checkpoint directory is given, the model’s default saver
 directory is used. Optionally appends current timestep to prevent overwriting previous
 checkpoint files. Turn off to be able to load model from the same given path argument as
 given here.

Parameters:
  • directory – Optional checkpoint directory.
  • append_timestep – Appends the current timestep to the checkpoint file if true.
Returns:

Checkpoint path were the model was saved.
tf_actions_and_internals(states, internals, update, deterministic)

Creates the TensorFlow operations for retrieving the actions (and posterior internal states) in reaction to the given input states (and prior internal states).

Parameters:
  • states – Dict of state tensors.
  • internals – List of prior internal state tensors.
  • update – Boolean tensor indicating whether this call happens during an update.
  • deterministic – Boolean tensor indicating whether action should be chosen
     deterministically.
Returns:

Actions and list of posterior internal state tensors.
tf_discounted_cumulative_reward(terminal, reward, discount, final_reward=0.0)

Creates the TensorFlow operations for calculating the discounted cumulative rewards for a given sequence of rewards.

Parameters:
  • terminal – Terminal boolean tensor.
  • reward – Reward tensor.
  • discount – Discount factor.
  • final_reward – Last reward value in the sequence.
Returns:

Discounted cumulative reward tensor.
tf_loss_per_instance(states, internals, actions, terminal, reward, update)

Creates the TensorFlow operations for calculating the loss per batch instance of the given input states and actions.

Parameters:
  • states – Dict of state tensors.
  • internals – List of prior internal state tensors.
  • actions – Dict of action tensors.
  • terminal – Terminal boolean tensor.
  • reward – Reward tensor.
  • update – Boolean tensor indicating whether this call happens during an update.
Returns:

Loss tensor.
tf_optimization(states, internals, actions, terminal, reward, update)

Creates the TensorFlow operations for performing an optimization update step based on the given input states and actions batch.

Parameters:
  • states – Dict of state tensors.
  • internals – List of prior internal state tensors.
  • actions – Dict of action tensors.
  • terminal – Terminal boolean tensor.
  • reward – Reward tensor.
  • update – Boolean tensor indicating whether this call happens during an update.
Returns:

The optimization operation.
tf_regularization_losses(states, internals, update)

Creates the TensorFlow operations for calculating the regularization losses for the given input states.

Parameters:
  • states – Dict of state tensors.
  • internals – List of prior internal state tensors.
  • update – Boolean tensor indicating whether this call happens during an update.
Returns:

Dict of regularization loss tensors.

MemoryAgent

class tensorforce.agents.MemoryAgent(states_spec, actions_spec, config)

Bases: tensorforce.agents.agent.Agent

The MemoryAgent class implements a replay memory, from which it samples batches to update the value function.

import_observations(observations)

Load an iterable of observation dicts into the replay memory.

Parameters:
  • observations – An iterable with each element containing an observation. Each
  • requires keys 'state','action','reward','terminal', 'internal'. (observation) –
  • an empty list [] for 'internal' if internal state is irrelevant. (Use) –

Returns:

BatchAgent

class tensorforce.agents.BatchAgent(states_spec, actions_spec, config)

Bases: tensorforce.agents.agent.Agent

The BatchAgent class implements a batch memory, which is cleared after every update.

Each agent requires the following Configuration parameters:

  • states: dict containing one or more state definitions.
  • actions: dict containing one or more action definitions.
  • preprocessing: dict or list containing state preprocessing configuration.
  • exploration: dict containing action exploration configuration.

The BatchAgent class additionally requires the following parameters:

  • batch_size: integer of the batch size.
  • keep_last_timestep: bool optionally keep the last observation for use in the next batch
observe(terminal, reward)

Adds an observation and performs an update if the necessary conditions are satisfied, i.e. if one batch of experience has been collected as defined by the batch size.

In particular, note that episode control happens outside of the agent since the agent should be agnostic to how the training data is created.

Parameters:
  • reward – float of a scalar reward
  • terminal – boolean whether episode is terminated or not

Returns: void

Deep-Q-Networks (DQN)

class tensorforce.agents.DQNAgent(states_spec, actions_spec, network_spec, config)

Bases: tensorforce.agents.memory_agent.MemoryAgent

Deep-Q-Network agent (DQN). The piece de resistance of deep reinforcement learning as described by Minh et al. (2015). Includes an option for double-DQN (DDQN; van Hasselt et al., 2015)

DQN chooses from one of a number of discrete actions by taking the maximum Q-value from the value function with one output neuron per available action. DQN uses a replay memory for experience playback.

Configuration:

Each agent requires the following configuration parameters:

  • states: dict containing one or more state definitions.
  • actions: dict containing one or more action definitions.
  • preprocessing: dict or list containing state preprocessing configuration.
  • exploration: dict containing action exploration configuration.

The MemoryAgent class additionally requires the following parameters:

  • batch_size: integer of the batch size.
  • memory_capacity: integer of maximum experiences to store.
  • memory: string indicating memory type (‘replay’ or ‘prioritized_replay’).
  • update_frequency: integer indicating the number of steps between model updates.
  • first_update: integer indicating the number of steps to pass before the first update.
  • repeat_update: integer indicating how often to repeat the model update.

Each model requires the following configuration parameters:

  • discount: float of discount factor (gamma).
  • learning_rate: float of learning rate (alpha).
  • optimizer: string of optimizer to use (e.g. ‘adam’).
  • device: string of tensorflow device name.
  • tf_summary: string directory to write tensorflow summaries. Default None
  • tf_summary_level: int indicating which tensorflow summaries to create.
  • tf_summary_interval: int number of calls to get_action until writing tensorflow summaries on update.
  • log_level: string containing logleve (e.g. ‘info’).
  • distributed: boolean indicating whether to use distributed tensorflow.
  • global_model: global model.
  • session: session to use.

The DQN agent expects the following additional configuration parameters:

  • target_update_frequency: int of states between updates of the target network.
  • update_target_weight: float of update target weight (tau parameter).
  • double_q_model: boolean indicating whether to use a double q-model.
  • clip_loss: float if not 0, uses the huber loss with clip_loss as the linear bound
  • scope: TensorFlow variable scope name (default: ‘vpg’)
  • batch_size: Positive integer (mandatory)
  • learning_rate: positive float (default: 1e-3)
  • discount: Positive float, at most 1.0 (default: 0.99)
  • normalize_rewards: Boolean (default: false)
  • entropy_regularization: None or positive float (default: none)
  • optimizer: Specification dict (default: Adam with learning rate 1e-3)
  • state_preprocessing: None or dict with (default: none)
  • exploration: None or dict with (default: none)
  • reward_preprocessing: None or dict with (default: none)
  • log_level: Logging level, one of the following values (default: ‘info’)
    • ‘info’, ‘debug’, ‘critical’, ‘warning’, ‘fatal’
  • summary_logdir: None or summary directory string (default: none)
  • summary_labels: List of summary labels to be reported, some possible values below (default: ‘total-loss’)
    • ‘total-loss’
    • ‘losses’
    • ‘variables’
    • ‘activations’
    • ‘relu’
  • summary_frequency: Positive integer (default: 1)

Normalized Advantage Functions

class tensorforce.agents.NAFAgent(states_spec, actions_spec, network_spec, config)

Bases: tensorforce.agents.memory_agent.MemoryAgent

NAF: https://arxiv.org/abs/1603.00748

  • scope: TensorFlow variable scope name (default: ‘vpg’)
  • batch_size: Positive integer (mandatory)
  • learning_rate: positive float (default: 1e-3)
  • discount: Positive float, at most 1.0 (default: 0.99)
  • normalize_rewards: Boolean (default: false)
  • entropy_regularization: None or positive float (default: none)
  • optimizer: Specification dict (default: Adam with learning rate 1e-3)
  • state_preprocessing: None or dict with (default: none)
  • exploration: None or dict with (default: none)
  • reward_preprocessing: None or dict with (default: none)
  • log_level: Logging level, one of the following values (default: ‘info’)
    • ‘info’, ‘debug’, ‘critical’, ‘warning’, ‘fatal’
  • summary_logdir: None or summary directory string (default: none)
  • summary_labels: List of summary labels to be reported, some possible values below (default: ‘total-loss’)
    • ‘total-loss’
    • ‘losses’
    • ‘variables’
    • ‘activations’
    • ‘relu’
  • summary_frequency: Positive integer (default: 1)

Deep-Q-learning from demonostration (DQFD)

class tensorforce.agents.DQFDAgent(states_spec, actions_spec, network_spec, config)

Bases: tensorforce.agents.memory_agent.MemoryAgent

Deep Q-learning from demonstration (DQFD) agent (Hester et al., 2017). This agent uses DQN to pre-train from demonstration data.

Configuration:

Each agent requires the following configuration parameters:

  • states: dict containing one or more state definitions.
  • actions: dict containing one or more action definitions.
  • preprocessing: dict or list containing state preprocessing configuration.
  • exploration: dict containing action exploration configuration.

Each model requires the following configuration parameters:

  • discount: float of discount factor (gamma).
  • learning_rate: float of learning rate (alpha).
  • optimizer: string of optimizer to use (e.g. ‘adam’).
  • device: string of tensorflow device name.
  • tf_summary: string directory to write tensorflow summaries. Default None
  • tf_summary_level: int indicating which tensorflow summaries to create.
  • tf_summary_interval: int number of calls to get_action until writing tensorflow summaries on update.
  • log_level: string containing logleve (e.g. ‘info’).
  • distributed: boolean indicating whether to use distributed tensorflow.
  • global_model: global model.
  • session: session to use.

The DQFDAgent class additionally requires the following parameters:

  • batch_size: integer of the batch size.

  • memory_capacity: integer of maximum experiences to store.

  • memory: string indicating memory type (‘replay’ or ‘prioritized_replay’).

  • min_replay_size: integer of minimum replay size before the first update.

  • update_rate: float of the update rate (e.g. 0.25 = every 4 steps).

  • target_network_update_rate: float of target network update rate (e.g. 0.01 = every 100 steps).

  • use_target_network: boolean indicating whether to use a target network.

  • update_repeat: integer of how many times to repeat an update.

  • update_target_weight: float of update target weight (tau parameter).

  • demo_sampling_ratio: float, ratio of expert data used at runtime to train from.

  • supervised_weight: float, weight of large margin classifier loss.

  • expert_margin: float of difference in Q-values between expert action and other actions enforced .. code-block:

    by the large margin function.

  • clip_loss: float if not 0, uses the huber loss with clip_loss as the linear bound

import_demonstrations(demonstrations)

Imports demonstrations, i.e. expert observations. Note that for large numbers of observations, set_demonstrations is more appropriate, which directly sets memory contents to an array an expects a different layout.

Parameters:demonstrations – List of observation dicts
observe(reward, terminal)

Adds observations, updates via sampling from memories according to update rate. DQFD samples from the online replay memory and the demo memory with the fractions controlled by a hyper parameter p called ‘expert sampling ratio.

Parameters:
  • reward –
  • terminal –
pretrain(steps)

Computes pretrain updates.

Parameters:steps – Number of updates to execute.
set_demonstrations(batch)

Set all demonstrations from batch data. Expects a dict wherein each value contains an array containing all states, actions, rewards, terminals and internals respectively.

Parameters:batch –

Vanilla Policy Gradient

class tensorforce.agents.VPGAgent(states_spec, actions_spec, network_spec, config)

Bases: tensorforce.agents.batch_agent.BatchAgent

Vanilla Policy Gradient agent as described by [Sutton et al. (1999)] (https://papers.nips.cc/paper/1713-policy-gradient-methods-for-reinforcement-learning-with-function-approximation.pdf).

  • scope: TensorFlow variable scope name (default: ‘vpg’)
  • batch_size: Positive integer (mandatory)
  • discount: Positive float, at most 1.0 (default: 0.99)
  • normalize_rewards: Boolean (default: false)
  • entropy_regularization: None or positive float (default: none)
  • gae_lambda: None or float between 0.0 and 1.0 (default: none)
  • optimizer: Specification dict (default: Adam with learning rate 1e-3)
  • baseline_mode: None, or one of ‘states’ or ‘network’ specifying the baseline input (default: none)
  • baseline: None or specification dict, or per-state specification for aggregated baseline (default: none)
  • baseline_optimizer: None or specification dict (default: none)
  • state_preprocessing: None or dict with (default: none)
  • exploration: None or dict with (default: none)
  • reward_preprocessing: None or dict with (default: none)
  • log_level: Logging level, one of the following values (default: ‘info’)
    • ‘info’, ‘debug’, ‘critical’, ‘warning’, ‘fatal’
  • summary_logdir: None or summary directory string (default: none)
  • summary_labels: List of summary labels to be reported, some possible values below (default: ‘total-loss’)
    • ‘total-loss’
    • ‘losses’
    • ‘variables’
    • ‘activations’
    • ‘relu’
  • summary_frequency: Positive integer (default: 1)

Trust Region Policy Optimization (TRPO)

class tensorforce.agents.TRPOAgent(states_spec, actions_spec, network_spec, config)

Bases: tensorforce.agents.batch_agent.BatchAgent

Trust Region Policy Optimization (Schulman et al., 2015) agent.

  • scope: TensorFlow variable scope name (default: ‘trpo’)
  • batch_size: Positive integer (mandatory)
  • learning_rate: Max KL divergence, positive float (default: 1e-2)
  • discount: Positive float, at most 1.0 (default: 0.99)
  • entropy_regularization: None or positive float (default: none)
  • gae_lambda: None or float between 0.0 and 1.0 (default: none)
  • normalize_rewards: Boolean (default: false)
  • likelihood_ratio_clipping: None or positive float (default: none)
  • baseline_mode: None, or one of ‘states’ or ‘network’ specifying the baseline input (default: none)
  • baseline: None or specification dict, or per-state specification for aggregated baseline (default: none)
  • baseline_optimizer: None or specification dict (default: none)
  • state_preprocessing: None or dict with (default: none)
  • exploration: None or dict with (default: none)
  • reward_preprocessing: None or dict with (default: none)
  • log_level: Logging level, one of the following values (default: ‘info’)
    • ‘info’, ‘debug’, ‘critical’, ‘warning’, ‘fatal’
  • summary_logdir: None or summary directory string (default: none)
  • summary_labels: List of summary labels to be reported, some possible values below (default: ‘total-loss’)
    • ‘total-loss’
    • ‘losses’
    • ‘variables’
    • ‘activations’
    • ‘relu’
  • summary_frequency: Positive integer (default: 1)

State preprocessing

The agent handles state preprocessing. A preprocessor takes the raw state input from the environment and modifies it (for instance, image resize, state concatenation, etc.). You can find information about our ready-to-use preprocessors here.

Building your own agent

If you want to build your own agent, it should always inherit from Agent. If your agent uses a replay memory, it should probably inherit from MemoryAgent, if it uses a batch replay that is emptied after each update, it should probably inherit from BatchAgent.

We distinguish between agents and models. The Agent class handles the interaction with the environment, such as state preprocessing, exploration and observation of rewards. The Model class handles the mathematical operations, such as building the tensorflow operations, calculating the desired action and updating (i.e. optimizing) the model weights.

To start building your own agent, please refer to this blogpost to gain a deeper understanding of the internals of the TensorForce library. Afterwards, have look on a sample implementation, e.g. the DQN Agent and DQN Model.