Pith. sign in

REVIEW 4 major objections 5 minor 38 references

CSAOT: Cooperative Multi-Agent System for Active Object Tracking

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that cooperative, role-based agents operating on a single device—combined with a Mixture-of-Policies mechanism—extend active object tracking duration and improve robustness, outperforming a single-agent baseline on…

desk verdict A coherent but weakly supported assembly of known components; the multi-agent benefit is not isolated, and the headline claims outrun the evidence. read the letter →

arxiv 2501.13994 v1 pith:7PMX6HGG submitted 2025-01-23 cs.CV cs.AIcs.RO

Hy Nguyen , Bao Pham , Hung Du , Srikanth Thudumu , Rajesh Vasa , Kon Mouzakis This is my paper · ORCID
classification cs.CVcs.AIcs.RO
keywords ActiveObjectTrackingMulti-AgentDeepReinforcementLearningMixtureofExpertsPoliciesDecentralizedTrainingSingle-deviceContinuousactionspaceObstacleavoidance
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 tries to establish that active object tracking—keeping a moving target in view by actively steering a camera or vehicle—can be improved by splitting the job among several specialized agents that all run on the same device, rather than by one monolithic agent or by adding extra helper hardware. The proposed system, CSAOT, assigns separate roles (detection, movement, obstacle, and final navigation) and gives each role its own reward signal, while a Mixture of Policies mechanism combines small expert policy networks per agent. The key empirical claim is that on the hardest test map, the multi-agent system sustains longer tracking episodes than a single-agent baseline, with average episode length 25±4 versus 18±5. If correct, the work suggests that cooperation among role-based agents on one device is a viable route to more robust and cheaper active tracking.

What carries the argument

The load-bearing mechanism is the role-based multi-agent decomposition combined with the Mixture of Policies (MoP). Four agents operate on one device from a unified observation (current frame, speed, acceleration, steering angle): the Detection Agent predicts a bounding box, the Movement Agent predicts the target center, the Obstacle Agent predicts the nearest obstacle distance, and the Navigation Agent takes all outputs plus image features from a pretrained residual image encoder to output acceleration and steering. Each agent's policy network is a Mixture of Policies—an adaptation of Mixture of Experts where a gating mechanism selects and reweights a small number of expert policy networks (four experts, top two selected) for each input—so the system can specialize per situation while keeping inference cheap. Task-specific rewards (Intersection-over-Union for detection, distance error for obstacles, Manhattan distance for movement) give each subtask a direct learning signal, and a recurrent memory module lets each agent retain past observations.

What would settle it

Train CSAOT and a SingleAgent baseline with matched parameter counts, observation inputs, and episode budgets on the Complex map; if the episode-length difference shrinks to the noise range, the claimed benefit of the multi-agent system is not supported.

Watch

Extended reading notes

Core claim

CSAOT is a hierarchical, fully decentralized multi-agent system for active object tracking in which three perception agents—bounding-box detection, target-center movement prediction, and obstacle-distance estimation—extract subtask information from the same camera frame, and a fourth navigation agent consumes those outputs plus the encoded image to produce acceleration and steering commands. All agents train with a clipped policy-gradient method under a Decentralized Training-Decentralized Execution scheme, each with a task-specific component reward in addition to the global tracking reward, and each agent's policy is a Mixture of Policies: a gating network selects and reweights a subset of small expert policy networks for the current observation. The paper claims this design improves robustness against occlusions and rapid motion, optimizes camera movement to extend tracking duration, and outperforms a SingleAgent baseline that shares the same architecture but lacks these adaptations, with the largest gap on the Complex map (25±4 vs 18±5 average episode length).

Load-bearing premise

The SingleAgent baseline is a valid control: it must match CSAOT in architecture capacity, training budget, and observation, differing only in the multi-agent role split and the novel reward and policy mechanisms.

Editorial extensions

If this is right

  • On complex maps with dynamic and static obstacles, CSAOT sustains longer average tracking episodes than a single-agent baseline (25±4 vs 18±5).
  • The framework keeps the multi-agent benefit on a single device, avoiding extra cameras or auxiliary hardware that prior multi-agent active-tracking approaches require.
  • Role-specific component rewards give first-layer agents a direct learning signal, avoiding the bottleneck of learning solely from a global reward.
  • MoP's gating over expert policies reduces inference cost enough for multiple agents to run in real time without sacrificing accuracy.
  • The same architecture generalizes to unseen maps, with the largest gains appearing where target paths are complicated and obstacle density is high.

Reading between the lines

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

  • If the result holds under matched capacity, the real source of improvement may be the auxiliary predictions (bounding box, center, obstacle) acting as a learned intermediate representation for the navigation agent, more than the multi-agent formulation per se; the paper does not isolate these two factors.
  • A testable extension is to ablate MoP and the component rewards separately within the multi-agent system to see which contribution drives the Complex-map gain, since the reported comparison bundles all novel adaptations together.
  • The SingleTurn cumulative-reward difference (−25.46 vs −31.37) hints that CSAOT also accumulates less penalty on trivial maps, but the paper does not analyze why; a follow-up could examine whether the component rewards shape behavior beyond episode length.
  • A further extension would test CSAOT against a multi-agent baseline with the same role split but no MoP, to determine whether the expert-policy gating or the role decomposition is the essential ingredient.
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.

Referee Report

4 major / 5 minor

Summary. This paper proposes CSAOT, a role-based multi-agent reinforcement learning system for active object tracking on a single device. The framework splits the task into four agents—detection, movement prediction, obstacle avoidance, and final navigation—each implemented with an LSTM encoder and a proposed Mixture of Policies (MoP) module, and trains them with Proximal Policy Optimization using a global tracking reward plus subtask-specific rewards. The authors evaluate in the AirSim simulator on four self-built maps (SingleTurn, SimpleLoop, SharpLoop, Complex), comparing episode length and cumulative reward against a hand-built SingleAgent baseline. The headline empirical claim is a gain in episode length on the Complex map (25±4 vs 18±5), interpreted as evidence that multi-agent role-based cooperation improves tracking performance.

Significance. The architecture is cleanly described, and the single-device, role-based decomposition is a plausible direction for reducing hardware cost in active tracking; the MoP mechanism is also a reasonable adaptation of sparse MoE to policy networks. If the reported gain were robust and attributable to the multi-agent decomposition, the paper would be a useful empirical contribution. However, the current evidence does not support that attribution: the SingleAgent baseline strips out the dense reward functions and MoP as well as the role split, the cumulative-reward metric is saturated at -150 on three of four maps, no seed-level statistics or significance tests are reported, and the authors' own qualitative description says the tracker loses the target after three turns on Complex and rarely recovers. The central claim therefore remains unverified rather than demonstrated.

major comments (4)
  1. [§5.2, Table 4; §4.2.1–4.2.2] The SingleAgent baseline is not a matched control for the multi-agent claim. Section 5.2 states that the baseline shares 'the same agent architecture but without any of our novel adaptations,' and those adaptations include the task-based component rewards in Eqs. (15)–(17) and the MoP policy in Eq. (18), in addition to the role decomposition itself. The Complex-map EL gain in Table 4 could therefore be due to reward shaping or to the extra policy-network capacity provided by MoP, not to role-based cooperation. This reading is supported by Section 6.4, which credits the component rewards for the performance improvement. Please add a single-agent ablation that retains the dense rewards and MoP but collapses the roles, and match parameter counts and training budgets across all conditions.
  2. [Table 4; §5.3] The quantitative support for the headline claim is statistically weak and partly floor-bound. No number of trials, episodes, or seeds is reported, no significance test is given, and on Complex the CSAOT EL of 25±4 overlaps with the SingleAgent EL of 18±5. Because Table 2 terminates episodes at CR=-150 and both methods reach -150 on Complex, SharpLoop, and SimpleLoop, CR is constant on three of four maps, leaving EL as the only discriminative metric, and EL itself is censored by the per-map maximum in Table 2. Please report per-seed results with confidence intervals or paired tests and use an evaluation protocol that is not saturated by the reward floor.
  3. [§5.3 vs. Abstract] The qualitative results in Section 5.3 directly undercut the abstract's claims of robustness and extended tracking duration. The text states that on Complex the tracker 'can only track within the first 3 turns, then loses track of the target' and that it 'can hardly get back on track' on most maps. There is no experiment involving occlusion or rapid target motion, and no count of recovery episodes. Please report first-loss time, post-loss tracking length, and recovery statistics, and evaluate explicit occlusion and rapid-motion scenarios before making robustness claims.
  4. [§5.2] The evaluation is only against a custom SingleAgent baseline; no existing AOT or MADRL method is compared, and all four maps are self-built with no external benchmark. Even if comparison requires adaptation, the paper should include at least one standard baseline (e.g., a PPO-based tracker or the Luo et al. [14] approach) so the reported EL values are calibrated and the single-device contribution is placed in context.
minor comments (5)
  1. [Table 1; Eqs. (9)–(17)] The reward-scaling coefficients λ_track, λ_nav, λ_diff, λ_detect, λ_obstacle, and λ_movement are never given values, so the reward functions cannot be reproduced from the paper.
  2. [§3.3, Eq. (3)] The sentence 'The remaining elements in S_K are reweighted for the corresponding expert's impact' appears to be a typo; the equation reweights the selected experts in S_K, not the remaining elements.
  3. [§5.2 and Table 4] The abbreviation 'ASOT' is used where 'AOT' is presumably intended, and the column header 'AR EL' appears to be a formatting artifact.
  4. [§3.1] Setting γ=1 makes the return an undiscounted sum; this should be justified because it interacts with the -150 floor and with the censored episode-length comparisons.
  5. [General reproducibility] The paper does not provide code, trained models, or map definitions, and the expert-network and LSTM sizes are unspecified; these omissions make independent replication difficult.

Circularity Check

1 steps flagged · score 2.0 of 10

Only exhibited circularity is the renaming of MoE as 'novel' MoP; the central multi-agent claim rests on a confounded but non-circular experiment.

  1. renaming known result [Sec. 1 (contributions); Sec. 3.3 Eq. (3); Sec. 4.2.2 Eq. (18)]
    "Introduced MoP, a novel mechanism to learn policy efficiently. ... The way w is calculated in this architecture is the same as the previously shown MoE method."

    Eq. (18) defines MoP as a weighted sum of selected sub-policy networks with renormalized weights; Eq. (3) defines MoE as a weighted sum of selected expert networks with renormalized weights. Substituting p_i(o; theta_i) for f_i(x; theta_i) leaves the formula unchanged, and the paper states the weight computation is 'the same as the previously shown MoE method.' The claimed novel mechanism therefore reduces to the MoE construction already introduced in the preliminaries, relabeled as Mixture of Policies. This is a contribution-labeling circularity, not a derivation of new capacity.

full rationale

This is an empirical systems paper. The central assertion that role-based multi-agent collaboration on a single device improves active tracking is supported by the Table 4 experiments, not by an analytic derivation, so the main derivation chain is not circular. The one definitional collapse I can exhibit is MoP: it is MoE by construction, as shown above. The CSAOT-versus-SingleAgent comparison is also not a matched control — the baseline omits the multi-agent decomposition, the component rewards (Eqs. 15-17), and MoP (Eq. 18) together — so the measured gain cannot isolate the multi-agent contribution; however, that is a control/validity problem, not a circularity, and per the hard rules it is not scored as such. Section 5.3 concedes that on Complex the tracker 'can only track within the first 3 turns' and that on most maps it can 'hardly get back on track,' which undercuts the abstract's robustness claims as a matter of evidence quality. The only self-citation, [17], supports a generic statement about DRL value functions and is not load-bearing. Score 2 reflects the single minor renaming step; the empirical claim has independent content and is not forced by construction.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central claim rests on the AirSim simulator as a faithful proxy, on the adequacy of the Dec-POMDP formulation with a shared observation, on the untested hypothesis that MoE gating improves continuous-control policies, on the assumption that per-agent subtask rewards do not conflict with the global objective, and on the validity of the SingleAgent baseline. None of these is demonstrated with independent evidence, and the reward scaling weights that would let a reader reproduce the system are not reported.

free parameters (8)
  • λ_track
    This reward weight is introduced in Eq. 9 and its value is never reported, even though it directly shapes the learned tracking policy.
  • λ_nav
    This reward weight is introduced in Eq. 11 and its value is never reported, even though it directly shapes the learned navigation policy.
  • λ_diff
    This reward weight is introduced in Eq. 14 and its value is never reported; it penalizes sudden acceleration changes.
  • λ_detect
    This reward weight is introduced in Eq. 15 and its value is never reported; it scales the detection agent's IoU reward.
  • λ_obstacle
    This reward weight is introduced in Eq. 16 and its value is never reported; it scales the obstacle agent's distance reward.
  • λ_movement
    This reward weight is introduced in Eq. 17 and its value is never reported; it scales the movement agent's center-distance reward.
  • MoP expert count and K = 4 experts, K=2
    Table 1; these architecture choices are set by hand without ablation or sensitivity analysis.
  • PPO training hyperparameters = lr=0.003, epsilon=0.99, epsilon_decay=0.9, epochs_per_sample=2
    Table 1; chosen by hand and not tuned systematically; the paper does not report how sensitive the results are to these values.
assumptions (7)
  • domain assumption The Dec-POMDP formulation of Eq. 1 with γ=1 and a shared observation source adequately models the AOT task.
    Section 3.1 sets the problem as a decentralized, partially observable, single-device cooperative task; this choice excludes other MADRL formulations and is not justified against them.
  • domain assumption AirSim is a faithful proxy for real-world active object tracking.
    Section 5.1 uses AirSim for all training and testing, while the conclusion claims real-world applicability; no sim-to-real validation is provided.
  • domain assumption Pre-trained ResNet50 features are sufficient for the detection, movement, and obstacle subtasks.
    Section 4.1.1 uses ResNet50 as the shared feature extractor without discussing fine-tuning or domain shift.
  • domain assumption PPO with continuous actions and DTDE reliably optimizes the multi-agent system.
    Section 3.4 assumes PPO's sample efficiency and continuous-action support transfer to this role-based setting; no comparison to other MADRL optimizers is made.
  • ad hoc to paper MoE gating improves policy diversity and robustness in continuous control.
    Section 4.2.2 states this as the hypothesis behind MoP, but no experiment isolates the effect of the gating mechanism.
  • ad hoc to paper Task-based component rewards accelerate learning without conflicting with the global tracking objective.
    Section 4.2.1 assumes per-agent rewards help; no theoretical or empirical analysis shows they do not distort the global objective.
  • ad hoc to paper The SingleAgent baseline is a valid control that isolates CSAOT's contributions.
    Section 5.2 claims the baseline 'shares the same agent architecture but without any of our novel adaptations,' but no architecture or capacity matching is shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CSAOT: Cooperative Multi-Agent System for Active Object Tracking." pith.science (2026). https://pith.science/paper/7PMX6HGG

@misc{pith2026250113994,
  author       = {Pith},
  title        = {Pith review of: CSAOT: Cooperative Multi-Agent System for Active Object Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PMX6HGG}},
  note         = {Machine review of arXiv:2501.13994}
}
read the original abstract

Object Tracking is essential for many computer vision applications, such as autonomous navigation, surveillance, and robotics. Unlike Passive Object Tracking (POT), which relies on static camera viewpoints to detect and track objects across consecutive frames, Active Object Tracking (AOT) requires a controller agent to actively adjust its viewpoint to maintain visual contact with a moving target in complex environments. Existing AOT solutions are predominantly single-agent-based, which struggle in dynamic and complex scenarios due to limited information gathering and processing capabilities, often resulting in suboptimal decision-making. Alleviating these limitations necessitates the development of a multi-agent system where different agents perform distinct roles and collaborate to enhance learning and robustness in dynamic and complex environments. Although some multi-agent approaches exist for AOT, they typically rely on external auxiliary agents, which require additional devices, making them costly. In contrast, we introduce the Collaborative System for Active Object Tracking (CSAOT), a method that leverages multi-agent deep reinforcement learning (MADRL) and a Mixture of Experts (MoE) framework to enable multiple agents to operate on a single device, thereby improving tracking performance and reducing costs. Our approach enhances robustness against occlusions and rapid motion while optimizing camera movements to extend tracking duration. We validated the effectiveness of CSAOT on various interactive maps with dynamic and stationary obstacles.

Figures

Figures reproduced from arXiv: 2501.13994 by the authors.

Figure 1
Figure 1. CSAOT framework: Each agent (Detection Agent, Obstacle Agent, Movement Agent) is responsible for different tasks [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Environment visualization, in clear weather mode [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Target path for testing (White line), from left to right: [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: EL during training with CSAOT in Complex map [4] Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. 2018. Counterfactual Multi-Agent Policy Gradients. Pro￾ceedings of the AAAI Conference on Artificial Intelligence 32, 1 (Apr.…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 18 canonical work pages

  1. [14]

    Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. 2019. End-to-end active object tracking and its real-world deployment via reinforcement learning. IEEE transactions on pattern analysis and machine intelligence 42, 6 (2019), 1317–1332

    work page 2019

  2. [1]

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. arXiv:1412.3555 [cs.NE] https://arxiv.org/abs/1412.3555

    arXiv 2014

  3. [2]

    Denzler and D

    J. Denzler and D. W. R. Paulus. 1994. Active motion detection and object tracking. In Proceedings of 1st International Conference on Image Processing

    work page 1994

  4. [3]

    Yannis Flet-Berliac, Johan Ferret, Olivier Pietquin, Philippe Preux, and Matthieu Geist. 2021. Adversarially Guided Actor-Critic. arXiv:2102.04376 [cs.LG] https: //arxiv.org/abs/2102.04376 Figure 5: EL during training with CSAOT in Complex map

    arXiv 2021

  5. [4]

    Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. 2018. Counterfactual Multi-Agent Policy Gradients. Pro- ceedings of the AAAI Conference on Artificial Intelligence 32, 1 (Apr. 2018). https://doi.org/10.1609/aaai.v32i1.11794

    doi:10.1609/aaai.v32i1.11794 2018

  6. [5]

    Foerster, Yannis M

    Jakob N. Foerster, Yannis M. Assael, Nando de Freitas, and Shimon Whiteson

  7. [6]

    Albert Gu and Tri Dao. 2024. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752 [cs.LG] https://arxiv.org/abs/2312.00752

    arXiv 2024

  8. [7]

    Albert Gu, Karan Goel, and Christopher Ré. 2022. Efficiently Modeling Long Sequences with Structured State Spaces. arXiv:2111.00396 [cs.LG] https://arxiv. org/abs/2111.00396

    arXiv 2022

Show all 38 references
  1. [8]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep Residual Learning for Image Recognition. arXiv:1512.03385 [cs.CV] https://arxiv.org/abs/ 1512.03385

    2015 arXiv

  2. [9]

    Mingxin Jiang, Tao Hai, Zhigeng Pan, Haiyan Wang, Yinjie Jia, and Chao Deng

  3. [10]

    Lefèvre and N

    S. Lefèvre and N. Vincent. 2004. Real time multiple object tracking based on active contours. In International Conference on Image Analysis and Recognition

    2004

  4. [11]

    W. Lei, H. Fu, and G. Sun. 2022. Active object tracking of free floating space manipulators based on deep reinforcement learning. Advances in Space Research (2022)

    2022

  5. [12]

    J. Li, J. Xu, F. Zhong, X. Kong, Y. Qiao, and Y. Wang. 2020. Pose-assisted multi- camera collaboration for active object tracking. In Proceedings of the AAAI Con- ference on Artificial Intelligence

    2020

  6. [13]

    Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. 2020. Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments. arXiv:1706.02275 [cs.LG] https://arxiv.org/abs/1706.02275

    2020 arXiv

  7. [15]

    J. Ma, Z. Zhao, X. Yi, J. Chen, and L. Hong. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

    2018

  8. [16]

    Rusu, Joel Veness, Marc G

    Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Kirkeby Fidjeland, Georg Ostrovski, Stig Petersen, Charlie Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wi...

  9. [17]

    Hy Nguyen, Srikanth Thudumu, Hung Du, Kon Mouzakis, and Rajesh Vasa

  10. [18]

    Zepeng Ning and Lihua Xie. 2024. A survey on multi-agent reinforcement learning and its application. Journal of Automation and Intelligence (2024)

    2024

  11. [19]

    Sindhu Padakandla. 2021. A survey of reinforcement learning algorithms for dynamically varying environments. ACM Computing Surveys (CSUR) 54, 6 (2021), 1–25

    2021

  12. [20]

    Tabish Rashid, Mikayel Samvelyan, Christian Schroeder de Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. 2018. QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning. arXiv:1803.11485 [cs.LG] https://arxiv.org/abs/1803.11485

    2018 arXiv

  13. [21]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  14. [22]

    Shazeer et al

    N. Shazeer et al. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)

    2017 arXiv

  15. [23]

    A. S. Silva, F. M. Q. Severgnini, and M. L. Oliveira. 2016. Object tracking by color and active contour models segmentation. IEEE Latin America Transactions (2016)

    2016

  16. [24]

    Maddison, Arthur Guez, L

    David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, L. Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Vedavyas Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy P. Lillicrap, Made...

    2016

  17. [25]

    Zahra Soleimanitaleb and Mohammad Ali Keyvanrad. 2022. Single Object Track- ing: A Survey of Methods, Datasets, and Evaluation Metrics.CoRR abs/2201.13066 (2022). arXiv:2201.13066 https://arxiv.org/abs/2201.13066

    2022 arXiv

  18. [26]

    Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. 2016. Learning Multiagent Communication with Backpropagation. arXiv:1605.07736 [cs.LG] https://arxiv. org/abs/1605.07736

    2016 arXiv

  19. [27]

    Leibo, Karl Tuyls, and Thore Graepel

    Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vini- cius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel. 2017. Value-Decomposition Networks For Cooperative Multi-Agent Learning. arXiv:1706.05296 [c...

    2017 arXiv

  20. [28]

    Lu Wang, Litong Fan, Long Zhang, Rongcheng Zou, and Zhen Wang. 2023. Syner- gistic effects of adaptive reward and reinforcement learning rules on cooperation. New Journal of Physics 25, 7 (2023), 073008

    2023

  21. [29]

    M. Xi, Y. Zhou, Z. Chen, and W. Zhou. 2021. Anti-distractor active object track- ing in 3D environments. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)

    2021

  22. [30]

    Chao Yu, Akash Velu, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu. 2022. The Surprising Effectiveness of PPO in Cooperative, Multi-Agent Games. arXiv:2103.01955 [cs.LG] https://arxiv.org/abs/2103.01955

    2022 arXiv

  23. [31]

    Zhang, L

    K. Zhang, L. Zhang, and M. H. Yang. 2013. Robust object tracking via active feature selection. IEEE Transactions on Image Processing (2013)

    2013

  24. [32]

    Tianle Zhang, Zhen Liu, Shiguang Wu, Zhiqiang Pu, and Jianqiang Yi. 2022. In- trinsic Reward with Peer Incentives for Cooperative Multi-Agent Reinforcement Learning. In 2022 International Joint Conference on Neural Networks (IJCNN) . 1–7. https://doi.org/10.1109/IJCNN55064.202...

    2022

  25. [33]

    Yihe Zhou, Shunyu Liu, Yunpeng Qing, Kaixuan Chen, Tongya Zheng, Yan- hao Huang, Jie Song, and Mingli Song. 2023. Is Centralized Training with Decentralized Execution Framework Centralized Enough for MARL? arXiv:2305.17352 [cs.AI] https://arxiv.org/abs/2305.17352

    2023 arXiv

  26. [2015]

    Nature 518 (2015), 529–533

    Human-level control through deep reinforcement learning. Nature 518 (2015), 529–533. https://api.semanticscholar.org/CorpusID:205242740

    2015

  27. [2016]

    arXiv:1605.06676 [cs.AI] https://arxiv.org/abs/1605.06676

    Learning to Communicate with Deep Multi-Agent Reinforcement Learning. arXiv:1605.06676 [cs.AI] https://arxiv.org/abs/1605.06676

    arXiv

  28. [2017]

    arXiv:1707.06347 [cs.LG] https://arxiv.org/abs/1707.06347

    Proximal Policy Optimization Algorithms. arXiv:1707.06347 [cs.LG] https://arxiv.org/abs/1707.06347

    arXiv

  29. [2019]

    IEEE Access 7 (2019), 32400–32407

    Multi-agent deep reinforcement learning for multi-object tracker. IEEE Access 7 (2019), 32400–32407

    2019

  30. [2023]

    Algorithms 16, 5 (2023), 227

    UAV dynamic object tracking with Lightweight Deep Vision Reinforcement Learning. Algorithms 16, 5 (2023), 227

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.