Growing with Experience: Growing Neural Networks in Deep Reinforcement Learning

Lukas Fehring; Marius Lindauer; Theresa Eimer

REVIEW 4 major objections 4 minor 29 references

Growing with Experience: Growing Neural Networks in Deep Reinforcement Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Growing a policy network's depth during PPO training gives agents solution rates that static deep networks cannot match.

desk verdict A plausible depth-growth idea whose headline numbers are likely confounded by BOHB's early-stopping; needs major experimental revision. read the letter →

arxiv 2506.11706 v1 pith:QYYZEWOT submitted 2025-06-13 cs.LG

Lukas Fehring , Marius Lindauer , Theresa Eimer This is my paper

classification cs.LG

keywords reinforcementlearningnetworkgrowthmorphismNet2NetPPOMiniHackMuJoCoAnthyperparameteroptimization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large networks in reinforcement learning often train poorly from scratch, so the policy capacity an agent can actually use stays small. This paper claims that growing the network's depth during training, rather than fixing its size, lets PPO agents have both trainability and capacity: a small network learns the fundamentals, and function-preserving insertions of new layers add expressive power as training proceeds. The method, GrowNN, uses Net2Net transformations at scheduled checkpoints so the policy is unchanged at the moment of growth. In the reported experiments, static deep networks barely solve MiniHack Room while growing networks reach solution rates up to 54 percent, and on MuJoCo Ant growing networks improve final reward by up to 72 percent relative to static networks of the same final size. If this holds, it provides a cheap way to scale RL policies without the training instability that normally accompanies deep networks.

What carries the argument

The load-bearing object is the Net2DeeperNet network morphism. It adds a layer of the same width as the layer below, initialized with no bias and with an idempotent activation such as ReLU; the added neuron at position $i$ copies the input of neuron $i$ to its output with weight one, and all other incoming weights are zero. Because the transformation satisfies $f(\theta)(s)=\theta(s)$ for every input $s$, the network's encoded function is unchanged at the moment of growth. The schedule that decides when to grow is equally central: growth happens after each fidelity of a modified BOHB procedure, so the same checkpoints used for early stopping double as capacity milestones.

What would settle it

Measure the policy's effective feature rank or action entropy at each growth checkpoint; if the reported gains still appear when policy complexity is flat or falling before a growth point, the monotonic-complexity assumption that motivates GrowNN is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that incremental depth growth converts network capacity from a source of training instability into an asset for PPO. GrowNN starts with a small feature extractor and, after each budget fidelity of the hyperparameter search, inserts an identity layer via the Net2DeeperNet network morphism: the new layer has no bias, uses an idempotent activation, and copies each input neuron to its corresponding output neuron while zeroing the others, so the input-output mapping is exactly preserved. Subsequent PPO updates then exploit the extra depth. Against static baselines of the same final size, the paper reports that this schedule lifts MiniHack Room solution rates from about 6 percent to up to 54 percent and improves Ant's final reward by up to 72 percent (65 percent in the body of the text). The intended contribution is an algorithm-agnostic growth strategy that recovers the benefit of deeper policies without retraining, distillation, or algorithmic redesign.

Load-bearing premise

The method assumes that policy complexity grows monotonically with the number of environment interactions, so a fixed schedule of depth increases at each fidelity checkpoint is helpful rather than harmful.

Editorial extensions

If this is right

PPO agents can be trained at final depths that static training cannot handle, because the initial small network stabilizes early learning and the identity-preserving insertions avoid the usual disruption of architecture changes.
The approach carries over to any RL algorithm that uses a differentiable policy network, because network morphisms do not depend on PPO's specific objective.
Because growth points are supplied by the hyperparameter optimization budget, GrowNN adds no new training phase or extra samples; the capacity increase happens inside the existing run.
The gains are a property of the training trajectory rather than the final architecture, since the grown network and the static baseline have the same final size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves untested whether the fixed fidelity schedule is optimal; shifting growth checkpoints earlier or later would reveal how much of the gain depends on timing rather than on growing per se.
The same identity-preserving mechanism should apply to widening layers or growing CNN channels, so the depth result is evidence that capacity-growth schedules in general help RL, not only deeper multi-layer perceptrons.
A direct measurement of policy complexity, such as effective feature rank or action entropy at each fidelity, would test the paper's rationale that capacity should arrive as the policy's complexity increases; the paper reports no such measure.
On Ant the static baseline's return stays near the avoid-death level while grown agents move forward, which suggests growth may help escape low-reward local optima; testing on sparse-reward environments would clarify whether this is a general property.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: a plausible idea, honestly situated in prior work, but the empirical headline is not yet supported. The likely confound is the BOHB pruning schedule, which can selectively kill static deep networks before they learn while letting GrowNN survive. That has to be addressed before the numbers mean anything.

The paper proposes GrowNN: start PPO with a small MLP, then insert identity layers via Net2DeeperNet at fixed interaction budgets, and tune the whole thing with a modified BOHB that uses a static fidelity schedule. The authors are upfront that incremental growth was used by Liu et al. (2024) and that Net2Net was used in Dota 2 (Berner et al.). The new part is the generic depth-growing schedule and the claim that it beats static networks of the same final size. I believe that finding is not in prior work.

What is genuinely good: the method is simple, the morphism preserves the policy exactly, and the related work is accurate rather than self-aggrandizing. The authors also explicitly say more evaluation is needed, which is a good sign.

The soft spots are serious, though. There is no code release, no seed count, no error bars, and no hyperparameter names or values, let alone the growth schedule. The number inconsistency matters: the abstract says "up to 48%" on MiniHack and "72%" on Ant; Section 4 says "over 50%" solution rate and "65%" final reward. Those are materially different claims. More importantly, the stress-test concern is real and load-bearing. BOHB's early stopping terminates runs that are poor at early fidelities. Static deep PPO networks often start slow and would look poor at those early checkpoints, so BOHB will preferentially prune exactly the static deep runs that might have improved. GrowNN starts small, survives pruning, and only grows later. Thus the comparison may reflect the survival schedule, not growth itself. The paper provides no full-budget unpruned runs for the static baselines, no ablation of schedule timing, and no measure of policy complexity. The central claim is not yet isolated.

This reads like a workshop paper with an interesting seed. I would not cite it as evidence yet, but I would bring it to a reading group if someone works on RL hyperparameter optimization or network growth. For peer review, I'd say send it out, but with clear expectation of major revision. The referee should demand full experimental details, seeds and error bars, unpruned baselines, and an ablation that separates growth from BOHB's early-stopping. If those come back positive, the result would be useful.

Verdict: conditional at best, but worth engaging.

Referee Report

4 major / 4 minor

Summary. The paper proposes GrowNN, a method for progressively increasing the depth of a neural network policy during deep reinforcement learning training. The method starts from a small network and periodically adds layers using Net2Net network-morphism transformations, which preserve the encoded function at the moment of growth. The authors integrate GrowNN into PPO and evaluate it on MiniHack Room 10x10 and MuJoCo Ant, comparing against static networks of the same final size. They report that GrowNN achieves higher solution rates and returns than static deep networks, with claimed improvements of up to 48% on MiniHack Room and 72% on Ant, and they argue that growth enables using larger networks that static training cannot handle.

Significance. If the claims are correct, GrowNN would be a simple and broadly applicable way to train larger RL policies without the instabilities of static deep networks, and it would connect network-morphism techniques with modern RL hyperparameter optimization. The method is easy to understand, the use of Net2Net transformations to preserve the policy during growth is principled, and the choice to grow the feature-extractor depth is well motivated. However, the experimental section as written does not yet establish the central claim: the reported numbers are internally inconsistent, no error bars or seed counts are given, and the comparison against static baselines is potentially confounded by the early-stopping behavior of the modified BOHB tuning procedure. The idea is promising, but the evidence is currently insufficient for a full journal publication.

major comments (4)

[Abstract and Section 4] The reported improvements are numerically inconsistent: the abstract states "up to 48% on MiniHack Room and 72% on Ant," the introduction states "solution rates from 6% to up to 54%," and Section 4 states "solution rates of over 50%" on MiniHack and "an improvement of 65% in terms of final reward" on Ant. The paper should state one consistent set of numbers, specify exactly which metric each percentage refers to (solution rate, final return, relative improvement), and report the corresponding standard errors or confidence intervals.
[Section 3 and Section 4] The comparison against static baselines is confounded by the modified BOHB early-stopping procedure. Section 3 explains that poorly performing runs are terminated after a predefined fidelity, and Section 4 reports that static deep networks achieve a 0% solution rate or do not improve over a return of 1000. Because static deep PPO networks typically learn slowly, BOHB may prune exactly those runs that would have improved later, while GrowNN survives early pruning because it starts small. The paper does not report full-budget, unpruned runs for the static baselines, nor the fraction of static runs that were terminated at each fidelity. Without such evidence, the reported failure of static networks cannot be attributed to an inherent inability to learn rather than to the pruning schedule.
[Section 4, Figure 2] The experimental details needed to reproduce and interpret the central comparison are missing. The paper does not report the tuned hyperparameter configurations, the exact growth schedule (number of fidelities, budget per fidelity, starting and final depth, and the layer-adding criterion), the number of random seeds, the evaluation protocol, or error bars/confidence intervals for the IQM curves in Figure 2. The claim that GrowNN outperforms static networks of the same final size depends on these details, and the current presentation does not rule out that the gains come from a favorable interaction with the BOHB budget schedule rather than from growth itself.
[Section 4] The paper states that baselines are "tuned with the same approach" but does not report the resulting hyperparameters or the budget used for tuning. Since the same BOHB procedure is used for both GrowNN and static baselines, the only difference between the two should be the growth mechanism. Without reporting the tuning outcome, the reader cannot determine whether the comparison is fair, e.g., whether the static baselines received their preferred hyperparameters or whether the GrowNN advantage is an artifact of the search trajectory.

minor comments (4)

[Section 3] The text says "the activation function needs to be idempotent (such as ReLu)"; the standard spelling is "ReLU." Also, the explanation would benefit from a brief note on why applying ReLU twice is equivalent to applying it once in the identity-initialized case.
[Figure 2] The figure legend and caption are confusing: for GrowNN networks the number of layers refers to the final size, but the plotted curves grow over time, and the reader cannot tell when the depth increases. The dotted vertical lines are not labeled with the corresponding depths, and the left panel includes a "Static (1 layer)" curve but no "GrowNN (1 layer)" curve; the caption or text should clarify the comparison structure.
[References] The reference for BOHB (Falkner et al., 2018) is listed as "Industrial applications of answer set programming," which is a different paper; the correct reference is Falkner, Klein, and Hutter, "BOHB: Robust and Efficient Hyperparameter Optimization at Scale," ICML 2018. Also, the reference list contains a broken entry "Proc. of ICLR'19icl" for DARTS and RL-DARTS.
[Throughout] There are minor wording and formatting issues: "Mujoco" should be "MuJoCo," "incrementally GrowNN-deeper networks" is awkward, and the introduction's phrase "from 6% to up to 54%" should be tied to a specific table or figure so that the reader can verify it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GrowNN's performance claims are empirical and not reducible to a fitted or self-cited input.

full rationale

GrowNN is an empirical method rather than a derivation. The core loop—start small, Net2Net-grow without changing the encoded function, evaluate on MiniHack and Ant—does not define any measured quantity in terms of another measured quantity. The paper's own description ('We start training a small network to learn an initial policy. Then we add layers without changing the encoded function') is a constructive algorithm, not a self-referential identity. The growth schedule is 'the number of environment interactions as a heuristic' with evenly spaced growing points, and the comparison is against static baselines 'tuned with the same approach'; no fitted parameter is renamed as a prediction. The self-citations—SMAC3 [Lindauer et al., 2022] for hyperparameter optimization and the AutoRL survey [Parker-Holder et al., 2022] for common MLP usage—are tooling and survey references. SMAC3 is an external, code-reproduced package that does not encode GrowNN's outcome, and the survey citation is not load-bearing for the central claim. Net2Net is an external transformation with an explicit function-preservation property, and no uniqueness theorem or prior-work ansatz is invoked to force the method. The BOHB early-stopping schedule could interact with static large networks (a selection confound), but that is a correctness or experimental-design concern, not a circular reduction: the reported solution rates are empirical outcomes, not outputs of the tuner by construction. The conclusion explicitly requests more evaluation, so the paper does not claim a closed-form derivation. Therefore no circular step can be quoted or exhibited, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The ledger is light on free parameters only because the paper reports almost none of its configuration; the growth schedule and PPO hyperparameters are tuned via BOHB but not specified. The main axioms are the standard Net2Net morphism property and the unmeasured monotonic-complexity assumption that motivates the schedule.

free parameters (2)

Growth schedule (fidelity count, budget per fidelity, starting and final depth) = Not reported
The schedule is selected per task via a modified BOHB procedure, but no concrete values are given; the reported gains depend on this schedule.
PPO hyperparameters = Not reported
Learning rate, batch size, and related settings are tuned with BOHB but not listed, making the comparison difficult to audit.

assumptions (3)

standard math Net2DeeperNet insertion with zero bias and idempotent activation preserves the encoded function (Chen et al., 2016)
The function-preserving property is taken from the cited Net2Net paper and not re-proven; Section 3 relies on it to justify that growth does not reset the policy.
ad hoc to paper Policy complexity increases monotonically with environment interactions, so a fixed schedule of depth increases is beneficial
Stated as the motivation in Section 3 ('small networks may learn simpler skills...larger network would be preferred') but never measured or ablated.
domain assumption BOHB hyperparameter configurations found for GrowNN also provide well-tuned static baselines under the same protocol
Section 4 asserts static baselines are tuned with the same approach, but no details or budgets are provided; if baselines are undertuned, the comparison is confounded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Growing with Experience: Growing Neural Networks in Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/QYYZEWOT

@misc{pith2026250611706,
  author       = {Pith},
  title        = {Pith review of: Growing with Experience: Growing Neural Networks in Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYYZEWOT}},
  note         = {Machine review of arXiv:2506.11706}
}

read the original abstract

While increasingly large models have revolutionized much of the machine learning landscape, training even mid-sized networks for Reinforcement Learning (RL) is still proving to be a struggle. This, however, severely limits the complexity of policies we are able to learn. To enable increased network capacity while maintaining network trainability, we propose GrowNN, a simple yet effective method that utilizes progressive network growth during training. We start training a small network to learn an initial policy. Then we add layers without changing the encoded function. Subsequent updates can utilize the added layers to learn a more expressive policy, adding capacity as the policy's complexity increases. GrowNN can be seamlessly integrated into most existing RL agents. Our experiments on MiniHack and Mujoco show improved agent performance, with incrementally GrowNN-deeper networks outperforming their respective static counterparts of the same size by up to 48% on MiniHack Room and 72% on Ant.

Figures

Figures reproduced from arXiv: 2506.11706 by the authors.

**Figure 2.** Training curves for static (solid lines) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Courville, Marc G

Max Schwarzer, Johan Samir Obando - Ceron, Aaron C. Courville, Marc G. Bellemare, Rishabh Agarwal, and Pablo Samuel Castro. Bigger, better, faster: Human-level atari with human-level efficiency. In Proc. of ICML '23 , 2023

work page 2023
[3]

Jha, and Asako Kanezaki

Kei Ota, Devesh K. Jha, and Asako Kanezaki. A framework for training larger networks for deep reinforcement learning. MLJ, 2024

work page 2024
[4]

Mixtures of experts unlock parameter scaling for deep RL

Johan Samir Obando - Ceron, Ghada Sokar, Timon Willi, Clare Lyle, Jesse Farebrother, Jakob Nicolaus Foerster, Gintare Karolina Dziugaite, Doina Precup, and Pablo Samuel Castro. Mixtures of experts unlock parameter scaling for deep RL . In Proc. of ICML '24 , 2024

work page 2024
[5]

T. Wei, C. Wang, Y. Rui, and C. W. Chen. Network morphism. In Proc. of ICML '16 , pages 564--572, 2016

work page 2016
[6]

T. Chen, I. Goodfellow, and J. Shlens. Net2net: Accelerating learning via knowledge transfer. In Proc. of ICLR '16 , 2016

work page 2016
[7]

Schulman, F

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv:1707.06347 [cs.LG], 2017

arXiv 2017
[8]

Berner, G

C. Berner, G. Brockman, B. Chan, V. Cheung, P. Debiak, C. Dennison, D. Farhi, Q. Fischer, S. Hashme, C. Hesse, R. J \' o zefowicz, S. Gray, C. Olsson, J. Pachocki, M. Petrov, H. Pond \' e de Oliveira Pinto, J. Raiman, T. Salimans, J. Schlatter, J. Schneider, S. Sidor, I. Sutskever, J. Tang, F. Wolski, and S. Zhang. Dota 2 with large scale deep reinforceme...

arXiv 1912

Show all 29 references

[9]

X. Wan, C. Lu, J. Parker - Holder, P. Ball, V. Nguyen, B. Ru, and M. Osborne. Bayesian generational population-based training. In Proc. of AutoML Conf'22 . PMLR, 2022

2022
[10]

J. Liu, J. Obando - Ceron, A. Courville, and L. Pan. Neuroplastic expansion in deep reinforcement learning. CoRR, abs/2410.07994, 2024

2024 arXiv
[11]

u ttler, Edward Grefenstette, and Tim Rockt \

Mikayel Samvelyan, Robert Kirk, Vitaly Kurin, Jack Parker - Holder, Minqi Jiang, Eric Hambro, Fabio Petroni, Heinrich K \" u ttler, Edward Grefenstette, and Tim Rockt \" a schel. Minihack the planet: A sandbox for open-ended reinforcement learning research. In Proc. of N eur I...

2021
[12]

Todorov, T

E. Todorov, T. Erez, and Y. Tassa. MuJoCo : A physics engine for model-based control. In Proc. of IROS '12 , pages 5026--5033, 2012

2012
[13]

Rumelhart, G

D. Rumelhart, G. Hinton, and R. Williams. Learning internal representations by error propagation. In Parallel Distributed Processing, volume 1, chapter 8, pages 318--362. MIT Press, 1985

1985
[14]

Parker-Holder, R

J. Parker-Holder, R. Rajan, X. Song, A. Biedenkapp, Y. Miao, T. Eimer, B. Zhang, V. Nguyen, R. Calandra, A. Faust, F. Hutter, and M. Lindauer. Automated reinforcement learning ( A uto RL ): A survey and open problems. Journal of Artificial Intelligence Research (JAIR), 74: 0 5...

2022
[15]

V. Mnih, K. Kavukcuoglu, D. Silver, A. Rusu, J. Veness, M. Bellemare, A. Graves, M. Riedmiller, A. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis. Human-level control through deep reinforcem...

2015
[16]

IMPALA: scalable distributed deep-rl with importance weighted actor-learner architectures

Lasse Espeholt, Hubert Soyer, R \' e mi Munos, Karen Simonyan, Volodymyr Mnih, Tom Ward, Yotam Doron, Vlad Firoiu, Tim Harley, Iain Dunning, Shane Legg, and Koray Kavukcuoglu. IMPALA: scalable distributed deep-rl with importance weighted actor-learner architectures. In Proc. o...

2018
[17]

K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proc. of CVPR '16 , pages 770--778, 2016

2016
[18]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In Proc. of N eur IPS '17 . Curran Associates, Inc., 2017

2017
[19]

A survey on transformers in reinforcement learning

Wenzhe Li, Hao Luo, Zichuan Lin, Chongjie Zhang, Zongqing Lu, and Deheng Ye. A survey on transformers in reinforcement learning. TMLR, 2023

2023
[20]

Elsken, J

T. Elsken, J. Metzen, and F. Hutter. Efficient multi-objective N eural A rchitecture S earch via lamarckian evolution. In Proc. of ICLR '19 iclr19
[21]

RL-DARTS: differentiable architecture search for reinforcement learning

Yingjie Miao, Xingyou Song, Daiyi Peng, Summer Yue, Eugene Brevdo, and Aleksandra Faust. RL-DARTS: differentiable architecture search for reinforcement learning. In Proc. of ICLR '22 , 2022

2022
[22]

H. Liu, K. Simonyan, and Y. Yang. DARTS : Differentiable architecture search. In Proc. of ICLR '19 iclr19
[23]

Gradmax: Growing neural networks using gradient information

Utku Evci, Bart van Merrienboer, Thomas Unterthiner, Fabian Pedregosa, and Max Vladymyrov. Gradmax: Growing neural networks using gradient information. In Proc. of ICLR '22 , 2022

2022
[24]

Autogrow: Automatic layer growing in deep convolutional networks

Wei Wen, Feng Yan, Yiran Chen, and Hai Li. Autogrow: Automatic layer growing in deep convolutional networks. In Proc. of KDD '20 , 2020

2020
[25]

Self expanding neural networks

Rupert Mitchell, Martin Mundt, and Kristian Kersting. Self expanding neural networks. arXiv:2307.04526 [cs.LG], 2024

2024 arXiv
[26]

Chau Pham, Piotr Teterwak, Soren Nelson, and Bryan A. Plummer. Mixturegrowth: Growing neural networks by recombining learned parameters. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV , 2024

2024
[27]

Falkner, G

A. Falkner, G. Friedrich, K. Schekotihin, R. Taupe, and E. Teppan. Industrial applications of answer set programming. KI-K \"u nstliche Intelligenz , pages 1--12, 2018

2018
[28]

Lindauer, K

M. Lindauer, K. Eggensperger, M. Feurer, A. Biedenkapp, D. Deng, C. Benjamins, T. Ruhkopf, R. Sass, and F. Hutter. SMAC3 : A versatile bayesian optimization package for H yperparameter O ptimization. JMLR, 23 0 (54): 0 1--9, 2022

2022
[29]

of ICLR '19 , 2019

Proc. of ICLR '19 , 2019

2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

Courville, Marc G

Max Schwarzer, Johan Samir Obando - Ceron, Aaron C. Courville, Marc G. Bellemare, Rishabh Agarwal, and Pablo Samuel Castro. Bigger, better, faster: Human-level atari with human-level efficiency. In Proc. of ICML '23 , 2023

work page 2023

[3] [3]

Jha, and Asako Kanezaki

Kei Ota, Devesh K. Jha, and Asako Kanezaki. A framework for training larger networks for deep reinforcement learning. MLJ, 2024

work page 2024

[4] [4]

Mixtures of experts unlock parameter scaling for deep RL

Johan Samir Obando - Ceron, Ghada Sokar, Timon Willi, Clare Lyle, Jesse Farebrother, Jakob Nicolaus Foerster, Gintare Karolina Dziugaite, Doina Precup, and Pablo Samuel Castro. Mixtures of experts unlock parameter scaling for deep RL . In Proc. of ICML '24 , 2024

work page 2024

[5] [5]

T. Wei, C. Wang, Y. Rui, and C. W. Chen. Network morphism. In Proc. of ICML '16 , pages 564--572, 2016

work page 2016

[6] [6]

T. Chen, I. Goodfellow, and J. Shlens. Net2net: Accelerating learning via knowledge transfer. In Proc. of ICLR '16 , 2016

work page 2016

[7] [7]

Schulman, F

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv:1707.06347 [cs.LG], 2017

arXiv 2017

[8] [8]

Berner, G

C. Berner, G. Brockman, B. Chan, V. Cheung, P. Debiak, C. Dennison, D. Farhi, Q. Fischer, S. Hashme, C. Hesse, R. J \' o zefowicz, S. Gray, C. Olsson, J. Pachocki, M. Petrov, H. Pond \' e de Oliveira Pinto, J. Raiman, T. Salimans, J. Schlatter, J. Schneider, S. Sidor, I. Sutskever, J. Tang, F. Wolski, and S. Zhang. Dota 2 with large scale deep reinforceme...

arXiv 1912