Advantage Actor Critic (A2C)#

A2C (synchronous version of A3C) is a model-free, stochastic on-policy policy gradient algorithm

Paper: Asynchronous Methods for Deep Reinforcement Learning


Algorithm#

Note

This algorithm implementation relies on the existence of parallel environments instead of parallel actor-learners


Algorithm implementation#

Main notation/symbols:
- policy function approximator (\(\pi_\theta\)), value function approximator (\(V_\phi\))
- states (\(s\)), actions (\(a\)), rewards (\(r\)), next states (\(s'\)), dones (\(d\))
- values (\(V\)), advantages (\(A\)), returns (\(R\))
- log probabilities (\(logp\))
- loss (\(L\))

Learning algorithm#


compute_gae(...)
def \(\;f_{GAE} (r, d, V, V_{_{last}}') \;\rightarrow\; R, A:\)
\(adv \leftarrow 0\)
\(A \leftarrow \text{zeros}(r)\)
# advantages computation
FOR each reverse iteration \(i\) up to the number of rows in \(r\) DO
IF \(i\) is not the last row of \(r\) THEN
\(V_i' = V_{i+1}\)
ELSE
\(V_i' \leftarrow V_{_{last}}'\)
\(adv \leftarrow r_i - V_i \, +\) discount_factor \(\neg d_i \; (V_i' \, -\) lambda \(adv)\)
\(A_i \leftarrow adv\)
# returns computation
\(R \leftarrow A + V\)
# normalize advantages
\(A \leftarrow \dfrac{A - \bar{A}}{A_\sigma + 10^{-8}}\)

_update(...)
# compute returns and advantages
\(V_{_{last}}' \leftarrow V_\phi(s')\)
\(R, A \leftarrow f_{GAE}(r, d, V, V_{_{last}}')\)
# sample mini-batches from memory
[[\(s, a, logp, V, R, A\)]] \(\leftarrow\) states, actions, log_prob, values, returns, advantages
# mini-batches loop
FOR each mini-batch [\(s, a, logp, V, R, A\)] up to mini_batches DO
\(logp' \leftarrow \pi_\theta(s, a)\)
# compute entropy loss
IF entropy computation is enabled THEN
\({L}_{entropy} \leftarrow \, -\) entropy_loss_scale \(\frac{1}{N} \sum_{i=1}^N \pi_{\theta_{entropy}}\)
ELSE
\({L}_{entropy} \leftarrow 0\)
# compute policy loss
\(L_{\pi_\theta} \leftarrow -\frac{1}{N} \sum_{i=1}^N A \; ratio\)
# compute value loss
\(V_{_{predicted}} \leftarrow V_\phi(s)\)
\(L_{V_\phi} \leftarrow \frac{1}{N} \sum_{i=1}^N (R - V_{_{predicted}})^2\)
# optimization step
reset \(\text{optimizer}_{\theta, \phi}\)
\(\nabla_{\theta, \, \phi} (L_{\pi_\theta} + {L}_{entropy} + L_{V_\phi})\)
\(\text{clip}(\lVert \nabla_{\theta, \, \phi} \rVert)\) with grad_norm_clip
step \(\text{optimizer}_{\theta, \phi}\)
# update learning rate
IF there is a learning_rate_scheduler THEN
step \(\text{scheduler}_{\theta} (\text{optimizer}_{\theta})\)
step \(\text{scheduler}_{\phi} (\text{optimizer}_{\phi})\)

Usage#

Note

Support for recurrent neural networks (RNN, LSTM, GRU and any other variant) is implemented in a separate file (a2c_rnn.py) to maintain the readability of the standard implementation (a2c.py)

# import the agent and its default configuration
from skrl.agents.torch.a2c import A2C, A2C_DEFAULT_CONFIG

# instantiate the agent's models
models = {}
models["policy"] = ...
models["value"] = ...  # only required during training

# adjust some configuration if necessary
cfg_agent = A2C_DEFAULT_CONFIG.copy()
cfg_agent["<KEY>"] = ...

# instantiate the agent
# (assuming a defined environment <env> and memory <memory>)
agent = A2C(models=models,
            memory=memory,  # only required during training
            cfg=cfg_agent,
            observation_space=env.observation_space,
            action_space=env.action_space,
            device=env.device)

Configuration and hyperparameters#

A2C_DEFAULT_CONFIG = {
    "rollouts": 16,                 # number of rollouts before updating
    "mini_batches": 1,              # number of mini batches to use for updating

    "discount_factor": 0.99,        # discount factor (gamma)
    "lambda": 0.95,                 # TD(lambda) coefficient (lam) for computing returns and advantages

    "learning_rate": 1e-3,                  # learning rate
    "learning_rate_scheduler": None,        # learning rate scheduler class (see torch.optim.lr_scheduler)
    "learning_rate_scheduler_kwargs": {},   # learning rate scheduler's kwargs (e.g. {"step_size": 1e-3})

    "state_preprocessor": None,             # state preprocessor class (see skrl.resources.preprocessors)
    "state_preprocessor_kwargs": {},        # state preprocessor's kwargs (e.g. {"size": env.observation_space})
    "value_preprocessor": None,             # value preprocessor class (see skrl.resources.preprocessors)
    "value_preprocessor_kwargs": {},        # value preprocessor's kwargs (e.g. {"size": 1})

    "random_timesteps": 0,          # random exploration steps
    "learning_starts": 0,           # learning starts after this many steps

    "grad_norm_clip": 0.5,          # clipping coefficient for the norm of the gradients

    "entropy_loss_scale": 0.0,      # entropy loss scaling factor

    "rewards_shaper": None,         # rewards shaping function: Callable(reward, timestep, timesteps) -> reward
    "time_limit_bootstrap": False,  # bootstrap at timeout termination (episode truncation)

    "experiment": {
        "directory": "",            # experiment's parent directory
        "experiment_name": "",      # experiment name
        "write_interval": 250,      # TensorBoard writing interval (timesteps)

        "checkpoint_interval": 1000,        # interval for checkpoints (timesteps)
        "store_separately": False,          # whether to store checkpoints separately

        "wandb": False,             # whether to use Weights & Biases
        "wandb_kwargs": {}          # wandb kwargs (see https://docs.wandb.ai/ref/python/init)
    }
}

Spaces#

The implementation supports the following Gym spaces / Gymnasium spaces

Gym/Gymnasium spaces

Observation

Action

Discrete

\(\square\)

\(\blacksquare\)

MultiDiscrete

\(\square\)

\(\blacksquare\)

Box

\(\blacksquare\)

\(\blacksquare\)

Dict

\(\blacksquare\)

\(\square\)

Models#

The implementation uses 1 stochastic (discrete or continuous) and 1 deterministic function approximator. These function approximators (models) must be collected in a dictionary and passed to the constructor of the class under the argument models

Notation

Concept

Key

Input shape

Output shape

Type

\(\pi_\theta(s)\)

Policy

"policy"

observation

action

Categorical /
Multi-Categorical /
Gaussian /
MultivariateGaussian

\(V_\phi(s)\)

Value

"value"

observation

1

Deterministic

Features#

Support for advanced features is described in the next table

Feature

Support and remarks

    pytorch    

    jax    

Shared model

for Policy and Value

\(\blacksquare\)

\(\square\)

RNN support

RNN, LSTM, GRU and any other variant

\(\blacksquare\)

\(\square\)

API (PyTorch)#

skrl.agents.torch.a2c.A2C_DEFAULT_CONFIG#

alias of {‘discount_factor’: 0.99, ‘entropy_loss_scale’: 0.0, ‘experiment’: {‘checkpoint_interval’: 1000, ‘directory’: ‘’, ‘experiment_name’: ‘’, ‘store_separately’: False, ‘wandb’: False, ‘wandb_kwargs’: {}, ‘write_interval’: 250}, ‘grad_norm_clip’: 0.5, ‘lambda’: 0.95, ‘learning_rate’: 0.001, ‘learning_rate_scheduler’: None, ‘learning_rate_scheduler_kwargs’: {}, ‘learning_starts’: 0, ‘mini_batches’: 1, ‘random_timesteps’: 0, ‘rewards_shaper’: None, ‘rollouts’: 16, ‘state_preprocessor’: None, ‘state_preprocessor_kwargs’: {}, ‘time_limit_bootstrap’: False, ‘value_preprocessor’: None, ‘value_preprocessor_kwargs’: {}}

class skrl.agents.torch.a2c.A2C(models: Mapping[str, Model], memory: Memory | Tuple[Memory] | None = None, observation_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, action_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, device: str | torch.device | None = None, cfg: dict | None = None)#

Bases: Agent

__init__(models: Mapping[str, Model], memory: Memory | Tuple[Memory] | None = None, observation_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, action_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, device: str | torch.device | None = None, cfg: dict | None = None) → None#

Advantage Actor Critic (A2C)

https://arxiv.org/abs/1602.01783

Parameters:

  • models (dictionary of skrl.models.torch.Model) – Models used by the agent

  • memory (skrl.memory.torch.Memory, list of skrl.memory.torch.Memory or None) – Memory to storage the transitions. If it is a tuple, the first element will be used for training and for the rest only the environment transitions will be added

  • observation_space (int, tuple or list of int, gym.Space, gymnasium.Space or None, optional) – Observation/state space or shape (default: None)

  • action_space (int, tuple or list of int, gym.Space, gymnasium.Space or None, optional) – Action space or shape (default: None)

  • device (str or torch.device, optional) – Device on which a tensor/array is or will be allocated (default: None). If None, the device will be either "cuda" if available or "cpu"

  • cfg (dict) – Configuration dictionary

Raises:

KeyError – If the models dictionary is missing a required key

_update(timestep: int, timesteps: int) → None#

Algorithm’s main update step

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

act(states: torch.Tensor, timestep: int, timesteps: int) → torch.Tensor#

Process the environment’s states to make a decision (actions) using the main policy

Parameters:

  • states (torch.Tensor) – Environment’s states

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

Returns:

Actions

Return type:

torch.Tensor

init(trainer_cfg: Mapping[str, Any] | None = None) → None#

Initialize the agent

post_interaction(timestep: int, timesteps: int) → None#

Callback called after the interaction with the environment

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

pre_interaction(timestep: int, timesteps: int) → None#

Callback called before the interaction with the environment

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

record_transition(states: torch.Tensor, actions: torch.Tensor, rewards: torch.Tensor, next_states: torch.Tensor, terminated: torch.Tensor, truncated: torch.Tensor, infos: Any, timestep: int, timesteps: int) → None#

Record an environment transition in memory

Parameters:

  • states (torch.Tensor) – Observations/states of the environment used to make the decision

  • actions (torch.Tensor) – Actions taken by the agent

  • rewards (torch.Tensor) – Instant rewards achieved by the current actions

  • next_states (torch.Tensor) – Next observations/states of the environment

  • terminated (torch.Tensor) – Signals to indicate that episodes have terminated

  • truncated (torch.Tensor) – Signals to indicate that episodes have been truncated

  • infos (Any type supported by the environment) – Additional information about the environment

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

class skrl.agents.torch.a2c.A2C_RNN(models: Mapping[str, Model], memory: Memory | Tuple[Memory] | None = None, observation_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, action_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, device: str | torch.device | None = None, cfg: dict | None = None)#

Bases: Agent

__init__(models: Mapping[str, Model], memory: Memory | Tuple[Memory] | None = None, observation_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, action_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, device: str | torch.device | None = None, cfg: dict | None = None) → None#

Advantage Actor Critic (A2C) with support for Recurrent Neural Networks (RNN, GRU, LSTM, etc.)

https://arxiv.org/abs/1602.01783

Parameters:

  • models (dictionary of skrl.models.torch.Model) – Models used by the agent

  • memory (skrl.memory.torch.Memory, list of skrl.memory.torch.Memory or None) – Memory to storage the transitions. If it is a tuple, the first element will be used for training and for the rest only the environment transitions will be added

  • observation_space (int, tuple or list of int, gym.Space, gymnasium.Space or None, optional) – Observation/state space or shape (default: None)

  • action_space (int, tuple or list of int, gym.Space, gymnasium.Space or None, optional) – Action space or shape (default: None)

  • device (str or torch.device, optional) – Device on which a tensor/array is or will be allocated (default: None). If None, the device will be either "cuda" if available or "cpu"

  • cfg (dict) – Configuration dictionary

Raises:

KeyError – If the models dictionary is missing a required key

_update(timestep: int, timesteps: int) → None#

Algorithm’s main update step

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

act(states: torch.Tensor, timestep: int, timesteps: int) → torch.Tensor#

Process the environment’s states to make a decision (actions) using the main policy

Parameters:

  • states (torch.Tensor) – Environment’s states

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

Returns:

Actions

Return type:

torch.Tensor

init(trainer_cfg: Mapping[str, Any] | None = None) → None#

Initialize the agent

post_interaction(timestep: int, timesteps: int) → None#

Callback called after the interaction with the environment

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

pre_interaction(timestep: int, timesteps: int) → None#

Callback called before the interaction with the environment

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

record_transition(states: torch.Tensor, actions: torch.Tensor, rewards: torch.Tensor, next_states: torch.Tensor, terminated: torch.Tensor, truncated: torch.Tensor, infos: Any, timestep: int, timesteps: int) → None#

Record an environment transition in memory

Parameters:

  • states (torch.Tensor) – Observations/states of the environment used to make the decision

  • actions (torch.Tensor) – Actions taken by the agent

  • rewards (torch.Tensor) – Instant rewards achieved by the current actions

  • next_states (torch.Tensor) – Next observations/states of the environment

  • terminated (torch.Tensor) – Signals to indicate that episodes have terminated

  • truncated (torch.Tensor) – Signals to indicate that episodes have been truncated

  • infos (Any type supported by the environment) – Additional information about the environment

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps


API (JAX)#

skrl.agents.jax.a2c.A2C_DEFAULT_CONFIG#

alias of {‘discount_factor’: 0.99, ‘entropy_loss_scale’: 0.0, ‘experiment’: {‘checkpoint_interval’: 1000, ‘directory’: ‘’, ‘experiment_name’: ‘’, ‘store_separately’: False, ‘wandb’: False, ‘wandb_kwargs’: {}, ‘write_interval’: 250}, ‘grad_norm_clip’: 0.5, ‘lambda’: 0.95, ‘learning_rate’: 0.001, ‘learning_rate_scheduler’: None, ‘learning_rate_scheduler_kwargs’: {}, ‘learning_starts’: 0, ‘mini_batches’: 1, ‘random_timesteps’: 0, ‘rewards_shaper’: None, ‘rollouts’: 16, ‘state_preprocessor’: None, ‘state_preprocessor_kwargs’: {}, ‘time_limit_bootstrap’: False, ‘value_preprocessor’: None, ‘value_preprocessor_kwargs’: {}}

class skrl.agents.jax.a2c.A2C(models: Mapping[str, Model], memory: Memory | Tuple[Memory] | None = None, observation_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, action_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, device: str | jax.Device | None = None, cfg: dict | None = None)#

Bases: Agent

__init__(models: Mapping[str, Model], memory: Memory | Tuple[Memory] | None = None, observation_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, action_space: int | Tuple[int] | gym.Space | gymnasium.Space | None = None, device: str | jax.Device | None = None, cfg: dict | None = None) → None#

Advantage Actor Critic (A2C)

https://arxiv.org/abs/1602.01783

Parameters:

  • models (dictionary of skrl.models.jax.Model) – Models used by the agent

  • memory (skrl.memory.jax.Memory, list of skrl.memory.jax.Memory or None) – Memory to storage the transitions. If it is a tuple, the first element will be used for training and for the rest only the environment transitions will be added

  • observation_space (int, tuple or list of int, gym.Space, gymnasium.Space or None, optional) – Observation/state space or shape (default: None)

  • action_space (int, tuple or list of int, gym.Space, gymnasium.Space or None, optional) – Action space or shape (default: None)

  • device (str or jax.Device, optional) – Device on which a tensor/array is or will be allocated (default: None). If None, the device will be either "cuda" if available or "cpu"

  • cfg (dict) – Configuration dictionary

Raises:

KeyError – If the models dictionary is missing a required key

_update(timestep: int, timesteps: int) → None#

Algorithm’s main update step

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

act(states: ndarray | jax.Array, timestep: int, timesteps: int) → ndarray | jax.Array#

Process the environment’s states to make a decision (actions) using the main policy

Parameters:

  • states (np.ndarray or jax.Array) – Environment’s states

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

Returns:

Actions

Return type:

np.ndarray or jax.Array

init(trainer_cfg: Mapping[str, Any] | None = None) → None#

Initialize the agent

post_interaction(timestep: int, timesteps: int) → None#

Callback called after the interaction with the environment

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

pre_interaction(timestep: int, timesteps: int) → None#

Callback called before the interaction with the environment

Parameters:

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps

record_transition(states: ndarray | jax.Array, actions: ndarray | jax.Array, rewards: ndarray | jax.Array, next_states: ndarray | jax.Array, terminated: ndarray | jax.Array, truncated: ndarray | jax.Array, infos: Any, timestep: int, timesteps: int) → None#

Record an environment transition in memory

Parameters:

  • states (np.ndarray or jax.Array) – Observations/states of the environment used to make the decision

  • actions (np.ndarray or jax.Array) – Actions taken by the agent

  • rewards (np.ndarray or jax.Array) – Instant rewards achieved by the current actions

  • next_states (np.ndarray or jax.Array) – Next observations/states of the environment

  • terminated (np.ndarray or jax.Array) – Signals to indicate that episodes have terminated

  • truncated (np.ndarray or jax.Array) – Signals to indicate that episodes have been truncated

  • infos (Any type supported by the environment) – Additional information about the environment

  • timestep (int) – Current timestep

  • timesteps (int) – Number of timesteps