R2D2¶

Overview¶

R2D2 was first proposed in Recurrent experience replay in distributed reinforcement learning. In RNN training with experience replay, the RL algorithms usually face the problem of representational drift and recurrent state staleness. R2D2 utilizes two approaches: stored states and burn-in to mitigate the aforementioned effects. R2D2 agent integrates these findings to achieve significant advances in the state of the art on Atari-57 and matches the state of the art on DMLab-30. The authors claim that, Recurrent Replay Distributed DQN (R2D2) is the first agent to achieve this using a single network architecture and fixed set of hyper-parameters.

Quick Facts¶

R2D2 is an off-policy, model-free and value-based RL algorithm,
R2D2 is essentially a DQN-based algorithm using a distributed framework, double Q networks, dueling architecture, n-step TD loss, and prioritized experience replay.
R2D2 now only supports discrete action spaces and uses eps-greedy for exploration same as DQN.
R2D2 uses the stored state and burn_in techniques to mitigate the effects of representational drift and recurrent state staleness.
The DI-engine implementation of R2D2 provides res_link key to support residual link in recurrent Q network.

Key Equations or Key Graphs¶

R2D2 agent is most similar to Ape-X, built upon prioritized distributed replay and n-step double Q-learning (with n = 5), generating experience by a large number of actors (typically 256) and learning from batches of replayed experience by a single learner. The Q network of R2D2 use the dueling network architecture and provide an LSTM layer after the convolutional stack.

Instead of regular \((s, a, r, s^')\) transition tuples, R2D2 stores fixed-length (m = 80) sequences of \((s, a, r)\) in replay, with adjacent sequences overlapping each other by 40 time steps, and never crossing episode boundaries. Specifically, the n-step targets used in R2D2 is:

Here, \(\theta^{-}\) denotes the target network parameters which are copied from the online network parameters \(\theta\) every 2500 learner steps.

R2D uses the mixture of max and mean absolute n-step TD-errors \(\delta_i\) as prioritization metrics for prioritized experience replay over the sequence:

Note

In our DI-engine implementation, at each unroll step, the input to the LSTM-based Q network is just observation and the last hidden state, excluding reward and one-hot action.

For more details about how to use RNN in DI-engine, users can refer to How to use RNN, for data arrangement process in R2D2, users can refer to the section data-arrangement, for the burn-in technique in R2D2, users can refer to the section burn-in-in-r2d2.

Extensions¶

R2D2 can be combined with:

Learning from demonstrations

Users can refer to R2D3 paper and R2D3 doc of our R2D3 implementation. R2D3 is an agent that makes efficient use of demonstrations to solve hard exploration problems in partially observable environments with highly variable initial conditions.

Transformers

Transformers-based agents take advantage of their powerful attention mechanism to learn better policies in those environments where long-term memory can be beneficial. Users can refer to GTrXL paper and r2d2_gtrxl doc of our GTrXL implementation.

Implementations¶

The default config of R2D2Policy is defined as follows:

class ding.policy.r2d2.R2D2Policy(cfg: EasyDict, model: Module | None = None, enable_field: List[str] | None = None)[source]

Overview:

Policy class of R2D2, from paper Recurrent Experience Replay in Distributed Reinforcement Learning . R2D2 proposes that several tricks should be used to improve upon DRQN, namely some recurrent experience replay tricks and the burn-in mechanism for off-policy training.

Config:

ID	Symbol	Type	Default Value	Description	Other(Shape)
1	`type`	str	r2d2	RL policy register name, refer to registry `POLICY_REGISTRY`	This arg is optional, a placeholder
2	`cuda`	bool	False	Whether to use cuda for network	This arg can be diff- erent from modes
3	`on_policy`	bool	False	Whether the RL algorithm is on-policy or off-policy
4	`priority`	bool	False	Whether use priority(PER)	Priority sample, update priority
5	`priority_IS` `_weight`	bool	False	Whether use Importance Sampling Weight to correct biased update. If True, priority must be True.
6	`discount_` `factor`	float	0.997, [0.95, 0.999]	Reward’s future discount factor, aka. gamma	May be 1 when sparse reward env
7	`nstep`	int	3, [3, 5]	N-step reward discount sum for target q_value estimation
8	`burnin_step`	int	2	The timestep of burnin operation, which is designed to RNN hidden state difference caused by off-policy
9	`learn.update` `per_collect`	int	1	How many updates(iterations) to train after collector’s one collection. Only valid in serial training	This args can be vary from envs. Bigger val means more off-policy
10	`learn.batch_` `size`	int	64	The number of samples of an iteration
11	`learn.learning` `_rate`	float	0.001	Gradient step length of an iteration.
12	`learn.value_` `rescale`	bool	True	Whether use value_rescale function for predicted value
13	`learn.target_` `update_freq`	int	100	Frequence of target network update.	Hard(assign) update
14	`learn.ignore_` `done`	bool	False	Whether ignore done for target value calculation.	Enable it for some fake termination env
15	`collect.n_sample`	int	[8, 128]	The number of training samples of a call of collector.	It varies from different envs
16	`collect.unroll` `_len`	int	1	unroll length of an iteration	In RNN, unroll_len>1

The network interface R2D2 used is defined as follows:

class ding.model.template.q_learning.DRQN(obs_shape: int | SequenceType, action_shape: int | SequenceType, encoder_hidden_size_list: SequenceType = [128, 128, 64], dueling: bool = True, head_hidden_size: int | None = None, head_layer_num: int = 1, lstm_type: str | None = 'normal', activation: Module | None = ReLU(), norm_type: str | None = None, res_link: bool = False)[source]

Overview:: The DRQN (Deep Recurrent Q-Network) is a neural network model combining DQN with RNN to handle sequential data and partially observable environments. It consists of three main components: encoder, rnn, and head. - Encoder: Extracts features from various observation inputs. - RNN: Processes sequential observations and other data. - Head: Computes Q-values for each action dimension.
Interfaces:: __init__, forward.

Note

The current implementation supports: - Two encoder types: FCEncoder and ConvEncoder. - Two head types: DiscreteHead and DuelingHead. - Three RNN types: normal (LSTM with LayerNorm), pytorch (PyTorch’s native LSTM), and gru. You can extend the model by customizing your own encoder, RNN, or head by inheriting this class.

forward(inputs: Dict, inference: bool = False, saved_state_timesteps: list | None = None) → Dict[source]

Overview:

Defines the forward pass of the DRQN model. Takes observation and previous RNN states as inputs and predicts Q-values.

Arguments:

inputs (Dict): Input data dictionary containing observation and previous RNN state.
inference (bool): If True, unrolls one timestep (used during evaluation). If False, unrolls the entire sequence (used during training).
saved_state_timesteps (Optional[list]): When inference is False, specifies the timesteps whose hidden states are saved and returned.

ArgumentsKeys:

obs (torch.Tensor): Raw observation tensor.
prev_state (list): Previous RNN state tensor, structure depends on lstm_type.

Returns:

outputs (Dict): The output of DRQN’s forward, including logit (q_value) and next state.

ReturnsKeys:

logit (torch.Tensor): Discrete Q-value output for each action dimension.
next_state (list): Next RNN state tensor.

Shapes:

obs (torch.Tensor): \((B, N)\) where B is batch size and N is obs_shape.
logit (torch.Tensor): \((B, M)\) where B is batch size and M is action_shape.

Examples:

>>> # Initialize input keys
>>> prev_state = [[torch.randn(1, 1, 64) for __ in range(2)] for _ in range(4)] # B=4
>>> obs = torch.randn(4,64)
>>> model = DRQN(64, 64) # arguments: 'obs_shape' and 'action_shape'
>>> outputs = model({'obs': inputs, 'prev_state': prev_state}, inference=True)
>>> # Validate output keys and shapes
>>> assert isinstance(outputs, dict)
>>> assert outputs['logit'].shape == (4, 64)
>>> assert len(outputs['next_state']) == 4
>>> assert all([len(t) == 2 for t in outputs['next_state']])
>>> assert all([t[0].shape == (1, 1, 64) for t in outputs['next_state']])

Benchmark¶

Benchmark and comparison of R2D2 algorithm¶
environment	best mean reward	config link
Pong (PongNoFrameskip-v4)	20	config_link_p
Qbert (QbertNoFrameskip-v4)	6000	config_link_q
SpaceInvaders (SpaceInvadersNoFrameskip-v4)	1400	config_link_s

References¶

Kapturowski S, Ostrovski G, Quan J, et al. Recurrent experience replay in distributed reinforcement learning[C]//International conference on learning representations. 2018.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, Martin Riedmiller: “Playing Atari with Deep Reinforcement Learning”, 2013; arXiv:1312.5602.
Schaul, T., Quan, J., Antonoglou, I., & Silver, D. (2015). Prioritized experience replay. arXiv preprint arXiv:1511.05952.
Van Hasselt, H., Guez, A., & Silver, D. (2016, March). Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence (Vol. 30, No. 1).
Wang, Z., Schaul, T., Hessel, M., Hasselt, H., Lanctot, M., & Freitas, N. (2016, June). Dueling network architectures for deep reinforcement learning. In International conference on machine learning (pp. 1995-2003). PMLR.
Horgan D, Quan J, Budden D, et al. Distributed prioritized experience replay[J]. arXiv preprint arXiv:1803.00933, 2018.

Other Public Implementations¶

seed_rl

ray