streaming-deep-rl 0.1.18

Streaming Deep Reinforcement Learning

Streaming Deep RL

Explorations into the proposed Streaming Deep Reinforcement Learning, from University of Alberta.

Once completed, if it checks out, will reach to integrate the Stream Q(λ) with Q-Transformer.

A recent testimony to Streaming AC(λ) variant can be found here. Will be incorporated into the repository as well with a few improvements.

Paper reading by Youtube AI/ML educator @hu-po.

The official repository can be found here.

Install

$ pip install streaming-deep-rl

Usage

import torch

from streaming_deep_rl import StreamingACLambda
from x_mlps_pytorch.normed_mlp import MLP

# actor and critic

actor = MLP(
    8, 128, 128, 128,
    norm_elementwise_affine = False,
    activate_last = True
)

critic = MLP(
    8, 128, 128,
    norm_elementwise_affine = False
)

# agent

agent = StreamingACLambda(
    actor = actor,
    critic = critic,
    dim_state = 8,
    dim_actor = 128,
    num_discrete_actions = 4,
    delay_steps = 7 # 7-step TD works well for me. next step TD tends to hit a performance wall
)

# get action from state and pass to environment or world model

state = torch.randn(8)
action, action_dist = agent(state, sample = True)

# environment or world model gives back

next_state = torch.randn(8)
reward = torch.tensor(1.)
done = torch.tensor(False)

# update at each timestep, "streaming"

agent.update(
    state = state,
    action = action,
    next_state = next_state,
    reward = reward,
    is_terminal = done
)

Lunar Lander

$ uv run train_lunar.py

Citations

@inproceedings{Elsayed2024StreamingDR,
    title   = {Streaming Deep Reinforcement Learning Finally Works},
    author  = {Mohamed Elsayed and Gautham Vasan and A. Rupam Mahmood},
    year    = {2024},
    url     = {https://api.semanticscholar.org/CorpusID:273482696}
}

@article{Nauman2024BiggerRO,
    title   = {Bigger, Regularized, Optimistic: scaling for compute and sample-efficient continuous control},
    author  = {Michal Nauman and Mateusz Ostaszewski and Krzysztof Jankowski and Piotr Milo's and Marek Cygan},
    journal = {ArXiv},
    year    = {2024},
    volume  = {abs/2405.16158},
    url     = {https://api.semanticscholar.org/CorpusID:270063045}
}

@misc{daley2025averagingnstepreturnsreduces,
    title   = {Averaging $n$-step Returns Reduces Variance in Reinforcement Learning}, 
    author  = {Brett Daley and Martha White and Marlos C. Machado},
    year    = {2025},
    eprint  = {2402.03903},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url     = {https://arxiv.org/abs/2402.03903}, 
}

@misc{chen2026cautiousweightdecay,
    title   = {Cautious Weight Decay}, 
    author  = {Lizhang Chen and Jonathan Li and Kaizhao Liang and Baiyu Su and Cong Xie and Nuo Wang Pierse and Chen Liang and Ni Lao and Qiang Liu},
    year    = {2026},
    eprint  = {2510.12402},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url     = {https://arxiv.org/abs/2510.12402}, 
}

@misc{kumar2024maintainingplasticitycontinuallearning,
    title   = {Maintaining Plasticity in Continual Learning via Regenerative Regularization}, 
    author  = {Saurabh Kumar and Henrik Marklund and Benjamin Van Roy},
    year    = {2024},
    eprint  = {2308.11958},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url     = {https://arxiv.org/abs/2308.11958}, 
}