Novel Multi-Agent Action Masked Deep Reinforcement Learning for General Industrial Assembly Lines Balancing Problems

Ali Mohamed Ali; Hashim A. Hashim; Luca Tirel

REVIEW 1 major objections 1 minor 46 references

Novel Multi-Agent Action Masked Deep Reinforcement Learning for General Industrial Assembly Lines Balancing Problems

T0 review · 1 major / 1 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that multi-agent PPO with action masking solves general industrial assembly-line balancing problems, reaching the authors' optimal-control baseline in roughly 800–1,050 episodes while cutting centralized PPO training time…

desk verdict The sequential feasibility check is a genuinely reusable idea, but the optimality claims rest on an unverified baseline from the authors' own prior DRL work, so the evaluation as written does not support the paper's headline claims. read the letter →

arxiv 2507.16635 v1 pith:Z5VMBQRR submitted 2025-07-22 cs.AI

Ali Mohamed Ali , Luca Tirel , Hashim A. Hashim This is my paper

classification cs.AI

keywords assemblylinebalancingreinforcementlearningmulti-agentactionmaskingProximalPolicyOptimizationMarkovDecisionProcessproductionschedulingdistributed

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

This paper seeks to show that a generic industrial assembly-line balancing problem, with no assumption about line type, can be modeled as a finite Markov Decision Process and solved by a decentralized multi-agent Proximal Policy Optimization algorithm. The paper claims that giving each workstation its own agent and filtering the agent's choices with a state-dependent action mask makes PPO converge to the same minimum finishing time as the authors' earlier baseline method in about 800–1,050 episodes, using roughly half the training time of centralized PPO and producing inference more than ten times faster than the baseline. It also claims the per-workstation decomposition changes the action-space growth from exponential to polynomial in the number of tasks. If true, the practical payoff is real-time re-scheduling: a trained network can output a near-optimal assignment in well under a second after a disruption, where an exact or model-based method would take seconds to minutes. The optimality claims all lean on the authors' prior baseline, which is the load-bearing reference.

What carries the argument

The two load-bearing mechanisms are the action mask and the sequential feasibility check. The mask is a state-dependent function $M(s)$ that maps the current factory state to a binary vector, with 0 marking infeasible assignments; for PPO the network's output probabilities on masked actions are set to negative infinity and renormalized with softmax, so only feasible actions have nonzero probability. The multi-agent version gives agent $i$ an action space of dimension $\delta A_i=\sum_{j=0}^{O_i}\binom{|J|}{j}$, one count per workstation, which is what turns the exponential centralized count $\delta A=2^{|I|\times|J|}$ into polynomial growth. When the per-agent actions conflict, Algorithm 1's sequential feasibility check feeds a fictitious copy of the factory state to agents in random order, updating the fictitious state and mask after each proposed action, and returns a feasible concatenated action; this is the mechanism that allows decentralized execution while preserving global constraint satisfaction.

What would settle it

Recompute the minimum finishing time for the paper's three test cases (3×5, 15×10, 10×15) with the same parameters from Tables I-III using an independent exact solver such as integer programming or exhaustive search; if the solver returns an ending time lower than the paper's [43] baseline for any case, then the multi-agent PPO convergence to 'optimal' and the robustness percentages are not established.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that decentralizing the assembly line by workstation—one agent per workstation, all observing the same full factory state—combined with a randomized sequential feasibility check, turns a hard combinatorial scheduling problem into a set of smaller learning problems without sacrificing solution quality. The same mask function used centrally is applied per agent, and when the independently proposed actions conflict, a fictitious environment replays them in random order so the final concatenated action is always feasible. The paper reports that the resulting multi-agent PPO reaches the baseline optimal ending time in 800, 820, and 1,050 episodes for the 3×5, 10×15, and 15×10 cases respectively, that the 3×5 training time drops from 00:21:26 (centralized PPO) to 00:11:14, and that inference at 0.36 seconds is more than ten times faster than the optimal-control baseline at 3.976 seconds. The reward function gives a completion bonus scaled by $1/(1+k^\alpha)$, so shorter schedules earn more, and the mask keeps every explored action feasible.

Load-bearing premise

The paper defines 'optimal' as the finishing time produced by the authors' own earlier method [43]; if that method is not the true minimum, the claims of optimal convergence, more than 90% optimality under random initialization, and tenfold inference speedups are all measured against an unverified benchmark.

Editorial extensions

If this is right

If the central claim is correct, trained multi-agent PPO agents can produce feasible schedules at inference times of about 0.36–0.42 seconds in the tested cases, compared with 3.976–45.436 seconds for the paper's optimal-control baseline, making same-day re-scheduling after disruptions feasible.
The action-space growth stays polynomial when the task count rises, so the decomposition is a more scalable route than centralized agents for larger assembly lines, although the paper only tests three configurations (3×5, 15×10, 10×15).
The 90%+ optimality rates over random initial states (95.53%, 92.11%, 91.87%) suggest the trained PPO policy generalizes to partially populated factory states without retraining, which is exactly what real factories face after breakdowns.
PPO outperforms DQN in the paper's comparison: PPO converges in hundreds of episodes while DQN needs thousands, and PPO uses smaller networks in the multi-agent setting.

Reading between the lines

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

Editorial inference: because the 'optimal' baseline is the authors' earlier method rather than an independent solver, the most informative next experiment is to rerun the same three cases with an exact integer programming solver; if the true optima are lower, the convergence and optimality percentages would need to be restated.
Editorial inference: the reported action-space reduction counts per-workstation combinations, but the sequential feasibility check still loops over all workstations, so the practical scaling bottleneck may move from action count to the cost of re-masking in that loop; a testable extension is measuring wall-clock time per episode as $|I|$ and $|J|$ grow.
Editorial inference: the paper does not test zero-shot transfer across different factory parameters; a natural follow-up is training on one set of durations, precedences, and capacities and evaluating on another to test the claimed generality of the MDP model.
Editorial inference: the reward is sparse, paid only when all tasks finish; one could add dense shaping terms without changing the optimum, which might accelerate convergence and make the method easier to tune for new line configurations.

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

The sequential feasibility check (SFC) is the one genuinely new piece here, and it's a reasonable mechanism for coordinating per-workstation agents. The MDP is a standard scheduling formulation, action masking is well-trodden, and multi-agent PPO/DQN is textbook; the contribution is the combination plus the SFC loop, and the action-space dimension counting is clear and honest. The writing is mostly readable: the environment, constraints, masks, and the SFC algorithm are all understandable. Credit is also due for testing both DQN and PPO and for reporting wall-clock training times, not just episode counts.

The soft spots are real and load-bearing. The 'optimal control (OC)' baseline in Sections VI and Table IV is reference [43], titled 'Action masked deep reinforcement learning for controlling industrial assembly lines' — a DRL paper, not an exact solver. Calling it an 'equivalent optimal control problem' mischaracterizes it. Unless [43] is verified to give provably optimal schedules, every 'convergence to k_opt' and 'over 90% optimal' claim is just convergence to that network's output. The stress-test note is correct: if [43] is a heuristic, the optimality results are unsupported.

There is also a numbers inconsistency: Section VII claims the multi-agent framework 'reduces training time by nearly tenfold,' but Table IV shows roughly a 4x reduction for the two larger cases and about 2x for the 3x5 case. That is not tenfold. Other issues are more standard: hyperparameters were tuned on the same three instances, only three small deterministic problems are tested, no code or data are provided (just 'available on request'), and the robustness test uses random initial states rather than new problem instances. A minor clarity issue is a broken sentence in Section III-C ('For the sake of simplicity, the maximal occupancy of workstation...').

This paper is for researchers working on DRL for scheduling who might reuse the SFC idea or the action-space analysis. It is a plausible engineering contribution, not a theoretical breakthrough. A serious referee could see value in the SFC architecture, but the evaluation needs a certified baseline (e.g., exact IP for small instances), code, and a corrected training-time claim.

Recommendation: send it to peer review, but with an explicit request to fix the baseline and broaden the evaluation. The core idea deserves the time; the current evidence doesn't yet back the claims.

Referee Report

1 major / 1 minor

Summary. The paper proposes a deep reinforcement learning (DRL) framework for general industrial assembly line balancing problems (ALBPs). It introduces an MDP model of a generic assembly line, an action-masking technique for both DQN and PPO, and a multi-agent decentralization in which each workstation is controlled by an individual agent, with a centralized training loop and a sequential feasibility check to guarantee feasible joint actions. The authors claim that the multi-agent action-masked PPO converges to optimal schedules, reduces training time by roughly 50% (and by nearly tenfold in a later statement) relative to centralized PPO, reduces the action-space growth from exponential to polynomial, and provides real-time inference more than ten times faster than an 'optimal control' baseline. Numerical experiments are reported for three instances (3x5, 15x10, 10x15 workstations/tasks) and a robustness test with random initial states.

Significance. If the claims were fully validated, the proposed framework would be a practical contribution to real-time assembly line scheduling: the MDP formulation is general, the action-masking and sequential-feasibility-check mechanisms are clearly described, and the paper provides pseudocode (Algorithms 1 and 2) and a detailed comparison of DQN and PPO in centralized and multi-agent variants. The empirical comparison of training times and inference times is also valuable. However, the central 'optimality' claims are currently measured against a baseline that is itself a learned DRL method from the authors' prior work, not against a certified exact solver; the generalization claim is tested only on instances used for hyperparameter tuning. These issues make the central contributions plausible but not yet established.

major comments (1)

[Section III-E, Eq. (8)] The reward function in Eq. (8) is non-zero only at the terminal step h[k]=1, making it extremely sparse. While the experiments show convergence, the paper does not discuss reward shaping, potential sparsity issues, or variance reduction techniques beyond action masking. A brief comment on how the sparse reward interacts with the PPO value estimation would strengthen the presentation.

minor comments (1)

[Section II-B, reference [23]] The description of reference [23] as combining RL with classical optimal control is not substantiated by the reference title ('Two-stage teaching-learning-based optimization method...'); the connection to RL is not obvious. Please either cite a more appropriate reference or elaborate on the relevance.

Circularity Check

1 steps flagged · score 4.0 of 10

All optimality claims in Sec. VI and Table IV are measured against [43], a self-cited DRL paper, so the 'optimal' label reduces to agreement with that baseline; the masking and multi-agent training-time contributions remain independent.

self citation load bearing [Section VI-A; Nomenclature (kopt); Fig. 4 caption; Table IV; Ref. [43]]
"The training processes of the agents were compared, with the solution to the problem formulated within the optimal control (OC) framework serving as a benchmark for performance evaluation. The agents were tuned to converge to the same outcomes as the equivalent OC problem."

The paper defines kopt via 'O.C.' and identifies OC with ref. [43], whose title is 'Action masked deep reinforcement learning for controlling industrial assembly lines' — a DRL method by two of the present authors, not an exact optimal-control or integer-programming oracle. Thus all convergence-to-optimal figures, >90% robustness percentages, and Table IV inference comparisons validate the agents against a prior learned model from the same group. The 'optimal' claim reduces, by benchmark construction, to agreeing with [43]'s solutions. This is load-bearing for the optimality conclusions, though Eq. (8), the mask constraints, and Eq. (9) remain independent.

full rationale

The core algorithmic derivation is self-contained: the reward function in Eq. (8) is defined directly from completion time, the action mask is computed from the feasibility constraints in the Appendix, and the multi-agent action-space dimension in Eq. (9) follows by combinatorial counting. None of these reduce to the baseline. The one load-bearing circular step is the evaluation: the paper's only ground truth for the 'optimal ending time' is kopt from the 'OC method' in [43], which is in fact a prior action-masked DRL paper by the same first two authors. All optimality and robustness claims in Sec. VI are therefore comparisons against a self-cited learned baseline rather than an externally certified optimum. Because the central training-time and polynomial-action-space contributions are independent and empirically demonstrated, the circularity is partial rather than total, so a score of 4 is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the reward parameters alpha/beta (not reported), hand-tuned network and algorithm hyperparameters, and the assumption that the authors' prior method [43] provides the optimal baseline. No new physical entities are introduced; the fictitious environment in the sequential feasibility check is a computational construct, not a new entity.

free parameters (6)

Reward exponent alpha = Not reported
Equation (8); controls the trade-off between early completion and reward magnitude; not specified.
Reward scale beta = Not reported
Equation (8); scales the terminal reward; not specified.
Learning rate = DQN: 1e-5, PPO: 3e-4
Table I; hand-tuned per experiment.
Hidden layer sizes = 534, 178, 258, 86
Table II; chosen empirically as 3-5x input size.
PPO learning frequency = 20
Table I; identified as the most crucial parameter for stability.
DQN target sync frequency = 10
Table I; critical for stability.

assumptions (5)

domain assumption The system is fully observable.
Section V-A states this assumption explicitly.
domain assumption The environment is deterministic.
Dynamics (1)-(3) are deterministic; no stochasticity is modeled.
standard math PPO and DQN converge to an optimal policy for the finite MDP given the hyperparameters.
The paper relies on standard convergence results for PPO [26] and DQN [27] without proof.
ad hoc to paper The optimal control method in [43] yields the true optimal schedule.
Used as the sole benchmark in Section VI; [43] is the authors' own prior work.
domain assumption Workstation capacities O_i are fixed and small, giving polynomial action-space growth.
The polynomial growth claim in Fig. 8(a) depends on O_i being constant; this condition is not stated in the claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Novel Multi-Agent Action Masked Deep Reinforcement Learning for General Industrial Assembly Lines Balancing Problems." pith.science (2026). https://pith.science/paper/Z5VMBQRR

@misc{pith2026250716635,
  author       = {Pith},
  title        = {Pith review of: Novel Multi-Agent Action Masked Deep Reinforcement Learning for General Industrial Assembly Lines Balancing Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z5VMBQRR}},
  note         = {Machine review of arXiv:2507.16635}
}

read the original abstract

Efficient planning of activities is essential for modern industrial assembly lines to uphold manufacturing standards, prevent project constraint violations, and achieve cost-effective operations. While exact solutions to such challenges can be obtained through Integer Programming (IP), the dependence of the search space on input parameters often makes IP computationally infeasible for large-scale scenarios. Heuristic methods, such as Genetic Algorithms, can also be applied, but they frequently produce suboptimal solutions in extensive cases. This paper introduces a novel mathematical model of a generic industrial assembly line formulated as a Markov Decision Process (MDP), without imposing assumptions on the type of assembly line a notable distinction from most existing models. The proposed model is employed to create a virtual environment for training Deep Reinforcement Learning (DRL) agents to optimize task and resource scheduling. To enhance the efficiency of agent training, the paper proposes two innovative tools. The first is an action-masking technique, which ensures the agent selects only feasible actions, thereby reducing training time. The second is a multi-agent approach, where each workstation is managed by an individual agent, as a result, the state and action spaces were reduced. A centralized training framework with decentralized execution is adopted, offering a scalable learning architecture for optimizing industrial assembly lines. This framework allows the agents to learn offline and subsequently provide real-time solutions during operations by leveraging a neural network that maps the current factory state to the optimal action. The effectiveness of the proposed scheme is validated through numerical simulations, demonstrating significantly faster convergence to the optimal solution compared to a comparable model-based approach.

Figures

Figures reproduced from arXiv: 2507.16635 by the authors.

**Figure 2.** The central agent architecture which introduces a modified version of the DQN/PPO agents making use of the action [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Action masking at the network output. this in mind, RL possesses a suitable structure for learner decentralization, thereby reducing computational complexity. In our scenario, the idea is to decentralize the assignment at the level of the workstations. We can consider each workstation as an agent aiming to solve a cooperative task. The constraints satisfaction plays an important role; in fact, logic must be implemen… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Action masking effect on the episodic length (finishing [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Sequential action feasibility check control logic. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Comparison of the episodic ending times (time neces [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 8.** Figure 8: Part (a) shows how the action space size increases, varying the number of tasks. Part (b) illustrates the episodic ending [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

[43]

Action masked deep reinforcement learning for controlling industrial assembly lines,

A. M. Ali and L. Tirel, “Action masked deep reinforcement learning for controlling industrial assembly lines,” in Proc. IEEE World AI IoT Congr . (AIIoT), 2023, pp. 0797–0803

work page 2023
[1]

Literature review of assembly line balancing problems,

P. Sivasankaran and P. Shahabudeen, “Literature review of assembly line balancing problems,” Int. J. Adv. Manuf. Technol. , vol. 73, pp. 1665– 1694, 2014

work page 2014
[2]

Literature review of industry 4.0 and related technologies,

E. Oztemel and S. Gursev, “Literature review of industry 4.0 and related technologies,” J. Intell. Manuf. , vol. 31, no. 1, pp. 127–182, 2020

work page 2020
[3]

The assembly-line balancing problem,

M. E. Salveson, “The assembly-line balancing problem,” Trans. ASME, vol. 77, no. 6, pp. 939–947, 1955

work page 1955
[4]

A survey on multi-agent reinforcement learning and its application,

Z. Ning and L. Xie, “A survey on multi-agent reinforcement learning and its application,” J. Autom. Intell. , 2024

work page 2024
[5]

Novel hybrid integrated pix2pix and wgan model with gradient penalty for binary images denoising,

L. Tirel, A. M. Ali, and H. A. Hashim, “Novel hybrid integrated pix2pix and wgan model with gradient penalty for binary images denoising,” Systems and Soft Computing , vol. 6, p. 200122, 2024

work page 2024
[6]

Segnet: A segmented deep learning based convolutional neural network approach for drones wildfire detection,

A. V . Jonnalagadda and H. A. Hashim, “Segnet: A segmented deep learning based convolutional neural network approach for drones wildfire detection,” Remote Sensing Applications: Society and Environment , vol. 34, p. 101181, 2024

work page 2024
[7]

Sampled-data control through model-free reinforcement learning with effective experience replay,

B. Xiao, H.-K. Lam, X. Su, Z. Wang, F. P.-W. Lo, S. Chen, and E. Yeatman, “Sampled-data control through model-free reinforcement learning with effective experience replay,” J. Autom. Intell., vol. 2, no. 1, pp. 20–30, 2023

work page 2023

Show all 46 references

[8]

Comprehensive and comparative analysis between transfer learning and custom built vgg and cnn-svm models for wildfire detection,

A. V . Jonnalagadda, H. A. Hashim, and A. Harris, “Comprehensive and comparative analysis between transfer learning and custom built vgg and cnn-svm models for wildfire detection,” in 2024 Sixth International Conference on Intelligent Computing in Data Sciences (ICDS) . IEEE, ...

2024
[9]

Reinforcement learning with soft temporal logic constraints using limit-deterministic generalized büchi automaton,

M. Cai, Z. Zhou, L. Li, S. Xiao, and Z. Kan, “Reinforcement learning with soft temporal logic constraints using limit-deterministic generalized büchi automaton,” J. Autom. Intell. , 2024

2024
[10]

A new optimal adaptive backstepping control ap- proach for nonlinear systems under deception attacks via reinforcement learning,

W. Chen and Q. Wei, “A new optimal adaptive backstepping control ap- proach for nonlinear systems under deception attacks via reinforcement learning,” J. Autom. Intell. , vol. 3, no. 1, pp. 34–39, 2024

2024
[11]

A comprehensive survey of robust deep learning in computer vision,

J. Liu and Y . Jin, “A comprehensive survey of robust deep learning in computer vision,” J. Autom. Intell. , 2023

2023
[12]

A taxonomy of line balancing problems and their solution approaches,

O. Battaïa and A. Dolgui, “A taxonomy of line balancing problems and their solution approaches,” Int. J. Prod. Econ. , vol. 142, no. 2, pp. 259– 277, 2013. 14

2013
[13]

Balancing of parallel u-shaped assembly lines,

I. Kucukkoc and D. Z. Zhang, “Balancing of parallel u-shaped assembly lines,” Comput. Oper . Res., vol. 64, pp. 233–244, 2015

2015
[14]

Scheduling and operator control in reconfigurable assembly systems,

D. Gyulai, B. Kádár, and L. Monostori, “Scheduling and operator control in reconfigurable assembly systems,” Procedia CIRP, vol. 63, pp. 459– 464, 2017

2017
[15]

Rolling horizon production scheduling of multi-model pcbs for several assembly lines,

J. Koskinen, C. Raduly-Baka, M. Johnsson, and O. S. Nevalainen, “Rolling horizon production scheduling of multi-model pcbs for several assembly lines,” Int. J. Prod. Res. , vol. 58, no. 4, pp. 1052–1073, 2020

2020
[16]

A genetic simulated annealing algorithm for parallel partial disassembly line balancing problem,

K. Wang, X. Li, L. Gao, P. Li, and S. M. Gupta, “A genetic simulated annealing algorithm for parallel partial disassembly line balancing problem,” Appl. Soft Comput. , vol. 107, p. 107404, 2021

2021
[17]

Mathematical model and bee algorithms for mixed-model assembly line balancing problem with physical human–robot collaboration,

Z. A. Çil, Z. Li, S. Mete, and E. Özceylan, “Mathematical model and bee algorithms for mixed-model assembly line balancing problem with physical human–robot collaboration,” Appl. Soft Comput. , vol. 93, p. 106394, 2020

2020
[18]

Designing assembly lines with humans and collaborative robots: A genetic approach,

M. Dalle Mura and G. Dini, “Designing assembly lines with humans and collaborative robots: A genetic approach,” CIRP Ann., vol. 68, no. 1, pp. 1–4, 2019

2019
[19]

A novel bi-level multi-objective genetic algorithm for integrated assembly line balancing and part feeding problem,

J. Chen, X. Jia, and Q. He, “A novel bi-level multi-objective genetic algorithm for integrated assembly line balancing and part feeding problem,” Int. J. Prod. Res. , vol. 61, no. 2, pp. 580–603, 2023

2023
[20]

Robust production planning and capacity control for flexible assembly lines,

D. Gyulai, B. Kádár, and L. Monosotori, “Robust production planning and capacity control for flexible assembly lines,” IF AC-PapersOnLine, vol. 48, no. 3, pp. 2312–2317, 2015

2015
[21]

An integrated framework for design, management and operation of reconfigurable assembly systems,

M. Manzini and et al, “An integrated framework for design, management and operation of reconfigurable assembly systems,” Omega, vol. 78, pp. 69–84, 2018

2018
[22]

Machine learning applications in production lines: A systematic literature review,

Z. Kang, C. Catal, and B. Tekinerdogan, “Machine learning applications in production lines: A systematic literature review,” Comput. Ind. Eng. , vol. 149, p. 106773, 2020

2020
[23]

Two-stage teaching-learning-based optimization method for flexible job-shop scheduling under machine breakdown,

R. Buddala and S. S. Mahapatra, “Two-stage teaching-learning-based optimization method for flexible job-shop scheduling under machine breakdown,” Int. J. Adv. Manuf. Technol., vol. 100, pp. 1419–1432, 2019

2019
[24]

Deep reinforcement learning-based dynamic scheduling in smart manufacturing,

L. Zhou, L. Zhang, and B. K. Horn, “Deep reinforcement learning-based dynamic scheduling in smart manufacturing,” Procedia CIRP , vol. 93, pp. 383–388, 2020

2020
[25]

Dynamic job-shop scheduling in smart manufacturing using deep reinforcement learning,

L. Wang, X. Hu, Y . Wang, S. Xu, S. Ma, K. Yang, Z. Liu, and W. Wang, “Dynamic job-shop scheduling in smart manufacturing using deep reinforcement learning,” Comput. Netw., vol. 190, p. 107969, 2021

2021
[26]

Proximal policy optimization algorithms,

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv:1707.06347, 2017

2017 arXiv
[27]

Human-level control through deep reinforcement learning,

V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015

2015
[29]

Safe reinforcement learning via shielding,

M. Alshiekh, R. Bloem, R. Ehlers, B. Könighofer, S. Niekum, and U. Topcu, “Safe reinforcement learning via shielding,” in Proc. AAAI Conf. Artif. Intell. , vol. 32, 2018

2018
[30]

A closer look at invalid action masking in policy gradient algorithms,

S. Huang and S. Ontañón, “A closer look at invalid action masking in policy gradient algorithms,” arXiv:2006.14171, 2020

2006 arXiv
[31]

Safe multi-agent reinforcement learning via shielding,

I. Elsayed-Aly, S. Bharadwaj, C. Amato, R. Ehlers, U. Topcu, and L. Feng, “Safe multi-agent reinforcement learning via shielding,” in Proc. AAMAS, Online, 2021

2021
[32]

Distributed actor–critic algo- rithms for multiagent reinforcement learning over directed graphs,

P. Dai, W. Yu, H. Wang, and S. Baldi, “Distributed actor–critic algo- rithms for multiagent reinforcement learning over directed graphs,”IEEE Trans. Neural Netw. Learn. Syst. , vol. 34, no. 10, pp. 7210–7221, 2022

2022
[33]

Multi-agent reinforce- ment learning: An overview,

L. Bu¸ soniu, R. Babuška, and B. D. Schutter, “Multi-agent reinforce- ment learning: An overview,” Innovations in Multi-Agent Systems and Applications, pp. 183–221, 2010

2010
[34]

A review of cooperative multi-agent deep reinforcement learning,

A. Oroojlooy and D. Hajinezhad, “A review of cooperative multi-agent deep reinforcement learning,” Appl. Intell. , vol. 53, no. 11, pp. 13 677– 13 722, 2023

2023
[35]

Cooperative multi-agent control using deep reinforcement learning,

J. K. Gupta, M. Egorov, and M. Kochenderfer, “Cooperative multi-agent control using deep reinforcement learning,” in Proc. AAMAS. Springer, 2017, pp. 66–83

2017
[36]

Weighted qmix: Expanding monotonic value function factorisation for deep multi-agent reinforcement learning,

T. Rashid, G. Farquhar, B. Peng, and S. Whiteson, “Weighted qmix: Expanding monotonic value function factorisation for deep multi-agent reinforcement learning,” Adv. Neural Inf. Process. Syst. , vol. 33, pp. 10 199–10 210, 2020

2020
[37]

Qplex: Duplex dueling multi-agent q-learning,

J. Wang, Z. Ren, T. Liu, Y . Yu, and C. Zhang, “Qplex: Duplex dueling multi-agent q-learning,” arXiv:2008.01062, 2020

2008 arXiv
[38]

Revisiting some common practices in cooperative multi-agent reinforcement learning,

W. Fu, C. Yu, Z. Xu, J. Yang, and Y . Wu, “Revisiting some common practices in cooperative multi-agent reinforcement learning,” in Proc. ICML, Baltimore, Maryland, USA, 2022

2022
[39]

The surprising effectiveness of ppo in cooperative multi-agent games,

C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of ppo in cooperative multi-agent games,” Adv. Neural Inf. Process. Syst. , vol. 35, pp. 24 611–24 624, 2022

2022
[40]

Facmac: Factored multi-agent centralised policy gradients,

B. Peng, T. Rashid, C. Schroeder de Witt, P.-A. Kamienny, P. Torr, W. Böhmer, and S. Whiteson, “Facmac: Factored multi-agent centralised policy gradients,” Adv. Neural Inf. Process. Syst. , vol. 34, pp. 12 208– 12 221, 2021

2021
[41]

Deep reinforcement learning for optimal planning of assembly line maintenance,

M. Geurtsen, I. Adan, and Z. Atan, “Deep reinforcement learning for optimal planning of assembly line maintenance,” J. Manuf. Syst., vol. 69, pp. 170–188, 2023

2023
[42]

Playing atari with deep reinforcement learn- ing,

V . Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wier- stra, and M. Riedmiller, “Playing atari with deep reinforcement learn- ing,” arXiv:1312.5602, 2013

2013 arXiv
[44]

A collaborative multi-agent deep reinforcement learning-based wireless power allocation with centralized training and decentralized execution,

A. Kopic, E. Perenda, and H. Gacanin, “A collaborative multi-agent deep reinforcement learning-based wireless power allocation with centralized training and decentralized execution,” IEEE Trans. Commun. , 2024

2024
[45]

Deep reinforcement learning for multi-agent power control in heterogeneous networks,

L. Zhang and Y .-C. Liang, “Deep reinforcement learning for multi-agent power control in heterogeneous networks,”IEEE Trans. Wirel. Commun., vol. 20, no. 4, pp. 2551–2564, 2020

2020
[46]

Age of information minimization using multi-agent uavs based on ai-enhanced mean field resource allocation,

Y . Emami, H. Gao, K. Li, L. Almeida, E. Tovar, and Z. Han, “Age of information minimization using multi-agent uavs based on ai-enhanced mean field resource allocation,” IEEE Trans. V eh. Technol., 2024

2024
[47]

Deep reinforcement learning for sim-to-real policy transfer of vtol-uavs offshore docking operations,

A. M. Ali, A. Gupta, and H. A. Hashim, “Deep reinforcement learning for sim-to-real policy transfer of vtol-uavs offshore docking operations,” Appl. Soft Comput. , p. 111843, 2024

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [43]

Action masked deep reinforcement learning for controlling industrial assembly lines,

A. M. Ali and L. Tirel, “Action masked deep reinforcement learning for controlling industrial assembly lines,” in Proc. IEEE World AI IoT Congr . (AIIoT), 2023, pp. 0797–0803

work page 2023

[2] [1]

Literature review of assembly line balancing problems,

P. Sivasankaran and P. Shahabudeen, “Literature review of assembly line balancing problems,” Int. J. Adv. Manuf. Technol. , vol. 73, pp. 1665– 1694, 2014

work page 2014

[3] [2]

Literature review of industry 4.0 and related technologies,

E. Oztemel and S. Gursev, “Literature review of industry 4.0 and related technologies,” J. Intell. Manuf. , vol. 31, no. 1, pp. 127–182, 2020

work page 2020

[4] [3]

The assembly-line balancing problem,

M. E. Salveson, “The assembly-line balancing problem,” Trans. ASME, vol. 77, no. 6, pp. 939–947, 1955

work page 1955

[5] [4]

A survey on multi-agent reinforcement learning and its application,

Z. Ning and L. Xie, “A survey on multi-agent reinforcement learning and its application,” J. Autom. Intell. , 2024

work page 2024

[6] [5]

Novel hybrid integrated pix2pix and wgan model with gradient penalty for binary images denoising,

L. Tirel, A. M. Ali, and H. A. Hashim, “Novel hybrid integrated pix2pix and wgan model with gradient penalty for binary images denoising,” Systems and Soft Computing , vol. 6, p. 200122, 2024

work page 2024

[7] [6]

Segnet: A segmented deep learning based convolutional neural network approach for drones wildfire detection,

A. V . Jonnalagadda and H. A. Hashim, “Segnet: A segmented deep learning based convolutional neural network approach for drones wildfire detection,” Remote Sensing Applications: Society and Environment , vol. 34, p. 101181, 2024

work page 2024

[8] [7]

Sampled-data control through model-free reinforcement learning with effective experience replay,

B. Xiao, H.-K. Lam, X. Su, Z. Wang, F. P.-W. Lo, S. Chen, and E. Yeatman, “Sampled-data control through model-free reinforcement learning with effective experience replay,” J. Autom. Intell., vol. 2, no. 1, pp. 20–30, 2023

work page 2023