Pith. sign in

REVIEW 3 major objections 6 minor 67 references

Towards Self-Evolving Agents: A Human-Inspired Adaptive Exploration-Exploitation Framework for Genetic Network Programming

T0 review · 3 major / 6 minor · reviewed 2026-07-15 · grok-4.5

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

Pith's one-line read A human-development schedule for GNP graphs improves agents by treating judgment transitions as thinking and processing nodes as action.

desk verdict Useful adaptive operators for GNP with a clean developmental framing and solid Tileworld stats; the generality claim is the soft spot, not the results themselves. read the letter →

arxiv 2607.11913 v1 pith:LKEHHR72 submitted 2026-07-04 cs.NE cs.AI

Ali Kohan , Mohamad Roshanzamir , Roohallah Alizadehsani , Seyedali Mirjalili This is my paper
classification cs.NEcs.AI
keywords GeneticNetworkProgrammingexploration-exploitationadaptivecrossovermutationcycleeliminationself-evolvingagentsTileworldagenticAI
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

Genetic Network Programming evolves directed graphs of judgment and processing nodes into interpretable agent strategies, yet most variants keep a fixed exploration–exploitation mix set by constant crossover and mutation rates. This paper argues that the mix should follow a developmental schedule: early generations should favor action (processing-node use and cycle-free exploration) the way children favor experimentation, while later generations should protect influential hubs and allow deeper deliberation the way adults do. The resulting Human-Inspired GNP (HGNP) therefore replaces uniform operators with three adaptive mechanisms—importance-weighted crossover that shields high in-degree nodes as iterations advance, a mutation that preferentially rewires judgment-to-judgment edges early on, and explicit elimination of processing and short judgment cycles—and supplies tunable parameters that let a designer dial the schedule to the size and difficulty of the search space. On the Tileworld benchmark the same schedule lifts standard GNP and two recent variants, with the combination HGNP-SBGNP producing the highest fitness, lowest variance, and most successful complete solutions. The claim is that a biologically motivated, time-varying bias is a more effective and more general control of the exploration–exploitation trade-off than ordinary probability tuning.

What carries the argument

Human-Inspired GNP (HGNP): an adaptive crossover probability Pc(n) that decreases with node in-degree and iteration count, a novel mutation PmNew(n) that severs judgment-to-judgment links early and relaxes later, and a cycle-elimination operator whose probability likewise decays with time; together they implement a child-to-adult exploration–exploitation trajectory inside the GNP graph.

What would settle it

Run the identical HGNP operators on a GNP domain whose optimal policy requires long early deliberation (e.g., a multi-step planning task with sparse rewards and many judgment nodes) and check whether the developmental schedule still outperforms carefully tuned fixed Pc/Pm; a clear reversal would falsify the claimed generality of the bias.

Watch

Extended reading notes

Core claim

Mapping GNP judgment-node transitions to deliberation and processing-node transitions to action, then scheduling adaptive crossover, novel mutation, and cycle elimination according to a normalized iteration clock, yields agents whose strategies converge faster and to higher fitness than those produced by fixed-rate genetic operators; the same schedule is portable across standard GNP and its recent variants and is most effective when the underlying search space is large.

Load-bearing premise

The paper assumes that treating judgment transitions as “thinking” and processing transitions as “action,” and then ramping protection and deliberation with iteration number, is the right inductive bias for GNP search spaces in general rather than an artifact of Tileworld’s sensor-action layout and the small program sizes tested.

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

3 major / 6 minor

Summary. The paper proposes Human-Inspired GNP (HGNP), an adaptive exploration–exploitation framework for Genetic Network Programming. Drawing an analogy between judgment-node transitions and deliberation and processing-node transitions and action, it introduces (i) an in-degree- and iteration-dependent crossover probability Pc(n) (Eqs. 1–2), (ii) an additional adaptive mutation that preferentially rewires judgment-to-judgment links early in search (Eqs. 4–5), and (iii) cycle-elimination operators for processing and judgment nodes (Eqs. 6–7). The operators are integrated with standard GNP, Simplified_GNP, and Situation-based GNP (SBGNP) and evaluated on three Tileworld instances with program sizes 3 and 5. Across 50 independent runs the authors report statistically significant fitness gains (Tables 3–8), with HGNP-SBGNP best overall, and argue that the framework is more effective than fixed Pc/Pm tuning and is applicable to almost all GNP variants.

Significance. If the reported gains hold and the operators transfer, the work supplies a concrete, low-overhead way to schedule exploration versus exploitation inside graph-structured evolutionary agents—an under-addressed issue in the GNP literature. Strengths include a public code repository, fifty independent runs, three environments of graded difficulty, two program sizes, t-tests with p-values, success counts, standard deviations, wall-clock times, and explicit ablation remarks (crossover alone can hurt; exploration operators alone plateau). The human-developmental framing is a novel inductive bias for GNP and is compatible with Industry 5.0 / XAI motivations. The contribution is primarily empirical and operator-level rather than a new theoretical guarantee; its lasting value hinges on whether the schedules remain useful outside Tileworld’s sensor–actuator node taxonomy.

major comments (3)
  1. Abstract and §4 claim that the judgment=deliberation / processing=action mapping plus the iteration-normalized schedules supply a general inductive bias for GNP search spaces, and that the modifications “can be applied to almost all GNP variants.” All experiments (Table 1, Fig. 12, Tables 3–8) use a single domain whose judgment functions are pure sensors and processing functions pure actuators, so the mapping is tautological for Tileworld. No second domain (continuous control, association-rule mining, multi-objective GNP, etc.) is reported. The statistical gains therefore establish that the operators help on Tileworld; they do not yet establish that the developmental schedule is the operative mechanism or that it transfers. Either add at least one qualitatively different benchmark or substantially narrow the generality claims in the abstract, introduction, and conclusion.
  2. §5.2–5.3 state that novel crossover alone produced no improvement (sometimes worse) and that the two exploration operators alone improved early iterations but then plateaued, while the joint design worked best. These remarks are useful but incomplete: there is no controlled ablation that isolates the iteration-dependent schedules (t, UB, αm, αelim) from the mere addition of extra mutation and cycle-breaking operators with fixed rates. Without that comparison it remains unclear whether the “child-to-adult” annealing is load-bearing or whether any extra early exploration plus late protection of high in-degree nodes would suffice. A factorial or schedule-vs-fixed ablation on at least one environment/program-size setting is needed to support the central mechanistic claim.
  3. Table 2 lists eight free parameters (Pc, Pm, PmNew, αm, Pelim, αelim, UB, and fitness weights w1–w3) that were “selected experimentally based on iterative testing.” §5.3 acknowledges hyperparameter sensitivity as a limitation, yet no sensitivity analysis, ranges, or selection protocol is reported. Because the paper’s selling point is that the adaptive framework is “more effective than tuning via crossover and mutation probabilities in standard GNP,” the reader needs evidence that the gains are robust to reasonable hyperparameter variation rather than the product of a carefully tuned schedule on this benchmark. A modest sensitivity study (or at least reporting the ranges tried) would make the comparison fair.
minor comments (6)
  1. Equation 3 asserts E[Pc(n)] = Pc for any UB ∈ [0,1]. The double integral is written with limits that treat Wi(n)+Wj(n) as ranging over [0,2] uniformly; a short derivation or numerical check would help readers verify the claim.
  2. Figure 13 panels are dense; adding a shared legend and marking the iteration at which HGNP-SBGNP overtakes baselines would improve readability.
  3. In Algorithm 1 the note about dividing Pelim by cycle length is important but easy to miss; consider elevating it into the main text of §4.2.2.
  4. Related-work coverage of adaptive GNP [42] is fair; a brief quantitative comparison (same Tileworld settings) would strengthen the claim of superiority over frequency-based adaptation.
  5. Minor typographical issues: “att ention” (abstract), “exploitatio n” (§2.3), inconsistent hyphenation of “exploration–exploitation,” and occasional missing spaces before citations.
  6. Code availability is welcome; stating the exact commit or release tag used for the reported tables would aid reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: HGNP operators are defined from iteration and in-degree, then scored on independent Tileworld fitness; self-citations supply only baselines.

full rationale

The paper defines adaptive Pc(n) from normalized node in-degrees Wi and iteration t (Eqs. 1–2), novel mutation PmNew(n) and cycle-elimination Pelim(n) from t (Eqs. 4–7), then measures success solely by the external Tileworld fitness of Eq. 8 (tiles dropped, residual steps, residual distances). That fitness is never an input to the schedules, so the reported gains (Tables 3–8) cannot be forced by construction. Self-citations to the authors’ prior SBGNP and Simplified_GNP papers appear only as comparison baselines whose search spaces are enlarged or reduced by HGNP; they do not underwrite the claimed improvement itself. The human-developmental mapping is an inductive bias, not a definitional identity that equates the operators to the fitness they later improve. Consequently the derivation chain is self-contained against an external benchmark and exhibits at most minor, non-load-bearing self-citation.

Assumptions & free parameters 8 free parameters · 4 assumptions · 1 invented entities

The central empirical claim rests on standard evolutionary-computation assumptions plus a handful of free schedule parameters and the novel human-development analogy. No new physical entities are postulated; the free parameters are the usual EA knobs plus the three HGNP-specific rates and their change rates.

free parameters (8)
  • Pc (base crossover rate) = 0.4
    Fixed at 0.4; used as the scale for the adaptive Pc(n).
  • Pm (traditional mutation rate) = 0.01
    Fixed at 0.01; retained alongside the novel mutation.
  • PmNew (novel mutation average rate) = 0.15
    Chosen experimentally; scales the early-iteration judgment-to-judgment mutation.
  • αm (novel-mutation change rate) = 0.15
    Controls how fast PmNew(n) decays with normalized iteration t.
  • Pelim (judgment-cycle elimination average rate) = 0.8
    Chosen experimentally; probability of breaking length-1 judgment cycles.
  • αelim (cycle-elimination change rate) = 0.5
    Controls decay of Pelim(n) with iteration.
  • UB (upper bound on normalized iteration for crossover) = 1
    Set to 1 so that the most influential nodes become completely protected by the final generation.
  • fitness weights w1,w2,w3 = 100,1,20
    Hand-set to 100,1,20 following prior Tileworld GNP papers; directly shape the ranking of solutions.
assumptions (4)
  • ad hoc to paper High in-degree nodes are the most influential and should be protected later in evolution.
    Introduced in §4.1 without formal proof; justified by informal graph-centrality intuition.
  • ad hoc to paper Judgment-node transitions correspond to deliberation and processing-node transitions to action, so early bias toward processing nodes mirrors child exploration.
    Core analogy of §4.2; maps human developmental literature onto GNP node types.
  • domain assumption Tileworld with the listed seven judgment and four processing functions is a representative benchmark for agent-control GNP.
    Standard in the GNP literature the authors cite; used as sole evaluation domain.
  • domain assumption Uniform random initialization and elite preservation of two individuals are adequate starting conditions.
    Inherited from classic GNP practice; not re-justified.
invented entities (1)
  • Human-Inspired GNP (HGNP) adaptive operators
    purpose: Dynamically schedule exploration (early processing bias + cycle breaking) versus exploitation (late protection of high-in-degree nodes).
    The three operators and the developmental schedule are the paper's primary technical contribution; they have no independent existence outside this work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Self-Evolving Agents: A Human-Inspired Adaptive Exploration-Exploitation Framework for Genetic Network Programming." pith.science (2026). https://pith.science/paper/LKEHHR72

@misc{pith2026260711913,
  author       = {Pith},
  title        = {Pith review of: Towards Self-Evolving Agents: A Human-Inspired Adaptive Exploration-Exploitation Framework for Genetic Network Programming},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LKEHHR72}},
  note         = {Machine review of arXiv:2607.11913}
}
read the original abstract

Recent advancements in agentic AI have increasingly moved toward graph-based methods, driven by the demand for explainable, human-centered, and non-linear reasoning workflows. A prominent example is Genetic Network Programming (GNP), a self-evolving algorithm that utilizes directed graphs to evolve interpretable decision structures for agents. As in most evolutionary algorithms, effectively balancing exploration and exploitation is a key aspect of GNP. However, this trade-off has received limited attention in the GNP literature. To address this gap, we draw inspiration from human developmental patterns, where children prioritize broad experimentation and action over deliberation, with this tendency reversing with age. By mapping transitions between GNP's judgment nodes to deliberation and processing nodes to action, we propose Human-Inspired GNP (HGNP), a novel adaptive framework that dynamically regulates the exploration-exploitation balance throughout the evolutionary process. The method consists of novel adaptive crossover and mutation operators, and a cycle elimination mechanism. HGNP not only improves the evolutionary process but also provides a framework for adjusting the exploration-exploitation balance based on the characteristics of the target environment and its search space. This approach is more effective than tuning via crossover and mutation probabilities in standard GNP. The modifications are general and can be applied to almost all GNP variants. When integrated with standard GNP and two recently introduced GNP variants and evaluated on the Tileworld benchmark, HGNP demonstrated significant performance improvement in agents' strategy. The combination of HGNP with Situation-based GNP (HGNP-SBGNP) achieved the best overall results.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 19 canonical work pages

  1. [1]

    The rise of agentic AI: implications, concerns, and the path forward,

    S. Murugesan, "The rise of agentic AI: implications, concerns, and the path forward," IEEE Intelligent Systems, vol. 40, no. 2, pp. 8-14, 2025

    2025

  2. [2]

    A survey of self -evolving agents: On path to artificial super intelligence,

    H.-a. Gao et al. , "A survey of self -evolving agents: On path to artificial super intelligence," arXiv preprint arXiv:2507.21046, vol. 1, 2025

    arXiv 2025

  3. [3]

    Agentic reasoning for large language models,

    T. Wei et al., "Agentic reasoning for large language models," arXiv preprint arXiv:2601.12538, 2026

    arXiv 2026

  4. [4]

    Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions,

    L. Longo et al., "Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions," Information Fusion, vol. 106, p. 102301, 2024

    2024

  5. [5]

    Human-Centered Pathways to Trustworthy AI in Healthcare: A Comparative Analysis of Explainable AI, Human-in-the-Loop, Hybrid AI, and Uncertainty Quantification Techniques,

    A. Kohan et al., "Human-Centered Pathways to Trustworthy AI in Healthcare: A Comparative Analysis of Explainable AI, Human-in-the-Loop, Hybrid AI, and Uncertainty Quantification Techniques," 2026

    2026

  6. [6]

    Application of explainable artificial intelligence (XAI) techniques in patients with intracranial hemorrhage: A systematic review,

    A. Kohan, A. Zahedi, R. Alizadehsani, R. S. Tan, and U. R. Acharya, "Application of explainable artificial intelligence (XAI) techniques in patients with intracranial hemorrhage: A systematic review," Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 15, no. 3, p. e70031, 2025

    2025

  7. [7]

    AgentGraph: Trace-to-Graph Platform for Interactive Analysis and Robustness Testing in Agentic AI Systems,

    Z. Wu et al., "AgentGraph: Trace-to-Graph Platform for Interactive Analysis and Robustness Testing in Agentic AI Systems," in Proceedings of the AAAI Conference on Artificial Intelligence, 2026, vol. 40, no. 48, pp. 41721-41723

    2026

  8. [8]

    Rethinking exploration –exploitation trade-off in reinforcement learning via cognitive consistency,

    D. Wang, W. Wei, L. Li, X. Wang, and J. Liang, "Rethinking exploration –exploitation trade-off in reinforcement learning via cognitive consistency," Neural Networks, vol. 187, p. 107342, 2025

    2025

Show all 67 references
  1. [9]

    Toward adaptive and coordinated transportation systems: A multi - personality multi-agent meta-reinforcement learning framework,

    S. Huang, C. Sun, R. -Q. Wang, and D. Pompili, "Toward adaptive and coordinated transportation systems: A multi - personality multi-agent meta-reinforcement learning framework," IEEE Transactions on Intelligent Transportation Systems, 2025

    2025

  2. [10]

    Agentevolver: Towards efficient self-evolving agent system,

    Y. Zhai et al., "Agentevolver: Towards efficient self-evolving agent system," arXiv preprint arXiv:2511.10395, 2025

    2025

  3. [11]

    Genetic programming and reinforcement learning on learning heuristics for dynamic scheduling: A preliminary comparison,

    M. Xu, Y. Mei, F. Zhang, and M. Zhang, "Genetic programming and reinforcement learning on learning heuristics for dynamic scheduling: A preliminary comparison," IEEE Computational Intelligence Magazine, vol. 19, no. 2, pp. 18-33, 2024

    2024

  4. [12]

    Comparison between Genetic Network Programming (GNP) and Genetic Programming (GP),

    K. Hirasawa, M. Okubo, H. Katagiri, J. Hu, and J. Murata, "Comparison between Genetic Network Programming (GNP) and Genetic Programming (GP)," in Proceedings of the 2001 Congress on Evolutionary Computation, 2001 2001, vol. 2, pp. 1276-1282 vol. 2, doi: 10.1109/CEC.2001.934337

    2001 doi

  5. [13]

    Enhancing Interpretability in Machine Learning: A Focus on Genetic Network Programming, Its Variants, and Applications,

    M. Roshanzamir, R. Alizadehsani, S. V. Moravvej, J. H. Joloudari, H. Alinejad-Rokny, and J. M. Gorriz, "Enhancing Interpretability in Machine Learning: A Focus on Genetic Network Programming, Its Variants, and Applications," in Artificial Intelligence for Neuroscience and Emot...

    2024

  6. [14]

    A novel estimation of distribution algorithm using graph-based chromosome representation and reinforcement learning,

    X. Li, B. Li, S. Mabu, and K. Hirasawa, "A novel estimation of distribution algorithm using graph-based chromosome representation and reinforcement learning," in IEEE Congress of Evolutionary Computation, 5-8 June 2011 2011, pp. 37-44, doi: 10.1109/CEC.2011.5949595

    2011 doi

  7. [15]

    Use of infeasible individuals in probabilistic model building genetic network programming,

    X. Li, S. Mabu, and K. Hirasawa, "Use of infeasible individuals in probabilistic model building genetic network programming," in 13th annual conference on Genetic and evolutionary computation , Dublin, Ireland, 2011, 2001659: ACM, pp. 601-608, doi: 10.1145/2001576.2001659

    2011 doi

  8. [16]

    An extended probabilistic model building genetic network programming using both of good and bad individuals,

    X. Li, S. Mabu, and K. Hirasawa, "An extended probabilistic model building genetic network programming using both of good and bad individuals," IEEJ Transactions on Electrical and Electronic Engineering, vol. 8, no. 4, pp. 339- 347, 2013, doi: 10.1002/tee.21864

    2013 doi

  9. [17]

    Genetic network programming with acquisition mechanisms of association rules,

    K. Shimada, "Genetic network programming with acquisition mechanisms of association rules," Journal of Advanced Computational Intelligence and Intelligent Informatics, vol. 10, no. 1, pp. 102 -111, 2006. [Online]. Available: http://ci.nii.ac.jp/naid/20000779642/en/

    2006

  10. [18]

    SemiSupervised Learning for Class Association Rule Mining Using Genetic Network Programming,

    S. Mabu, T. Higuchi, and T. Kuremoto, "SemiSupervised Learning for Class Association Rule Mining Using Genetic Network Programming," IEEJ Transactions on Electrical and Electronic Engineering, vol. 15, no. 5, pp. 733-740, 2020, doi: https://doi.org/10.1002/tee.23109

    2020 doi

  11. [19]

    Attribute Selection Based Genetic Network Programming for Intrusion Detection System,

    Y. Xu, Y. Sun, Z. Ma, H. Zhao, Y. Wang, and N. Lu, "Attribute Selection Based Genetic Network Programming for Intrusion Detection System," Journal of Advanced Computational Intelligence and Intelligent Informatics, vol. 26, no. 5, pp. 671-683, 2022, doi: 10.20965/jaciii.2022.p0671

    2022 doi

  12. [20]

    An Intrusion -Detection Model Based on Fuzzy Class - Association-Rule Mining Using Genetic Network Programming,

    S. Mabu, C. Chen, N. Lu, K. Shimada, and K. Hirasawa, "An Intrusion -Detection Model Based on Fuzzy Class - Association-Rule Mining Using Genetic Network Programming," Transactions on Systems, Man, and Cybernetics, vol. 41, no. 1, pp. 130-139, 2011, doi: 10.1109/tsmcc.2010.2050685

    2011 doi

  13. [21]

    Integrated fuzzy GNP rule mining with distance-based classification for intrusion detection system,

    N. Lu, S. Mabu, T. Wang, and K. Hirasawa, "Integrated fuzzy GNP rule mining with distance-based classification for intrusion detection system," in IEEE International Conference on Systems, Man, and Cybernetics, 14-17 Oct. 2012 2012, pp. 1569-1574, doi: 10.1109/ICSMC.2012.6377960

    2012 doi

  14. [22]

    Industry 5.0—A human-centric solution,

    S. Nahavandi, "Industry 5.0—A human-centric solution," Sustainability, vol. 11, no. 16, p. 4371, 2019

    2019

  15. [23]

    From Industry 4.0 digital manufacturing to Industry 5.0 digital society: a roadmap toward human -centric, sustainable, and resilient production,

    M. Ghobakhloo, H. A. Mahdiraji, M. Iranmanesh, and V. Jafari-Sadeghi, "From Industry 4.0 digital manufacturing to Industry 5.0 digital society: a roadmap toward human -centric, sustainable, and resilient production," Information Systems Frontiers, pp. 1-33, 2024

    2024

  16. [24]

    State of Industry 5.0 —Analysis and identification of current research trends,

    A. Akundi, D. Euresti, S. Luna, W. Ankobiah, A. Lopes, and I. Edinbarough, "State of Industry 5.0 —Analysis and identification of current research trends," Applied System Innovation, vol. 5, no. 1, p. 27, 2022

    2022

  17. [25]

    Industry 4.0 and Industry 5.0—Inception, conception and perception,

    X. Xu, Y. Lu, B. Vogel-Heuser, and L. Wang, "Industry 4.0 and Industry 5.0—Inception, conception and perception," Journal of manufacturing systems, vol. 61, pp. 530-535, 2021

    2021

  18. [26]

    Human -machine interaction towards Industry 5.0: Human -centric smart manufacturing,

    J. Yang, Y. Liu, and P. L. Morgan, "Human -machine interaction towards Industry 5.0: Human -centric smart manufacturing," Digital Engineering, p. 100013, 2024

    2024

  19. [27]

    Genetic network programming with parallel processing for association rule mining in large and dense databases,

    E. Gonzales, K. Shimada, S. Mabu, K. Hirasawa, and J. Hu, "Genetic network programming with parallel processing for association rule mining in large and dense databases," in Proceedings of the 9th annual conference on Genetic and evolutionary computation, 2007, pp. 1512-1512

    2007

  20. [28]

    Automatic Calibration of Piezoelectric Bed-Leaving Sensor Signals Using Genetic Network Programming Algorithms,

    H. Madokoro, S. Nix, and K. Sato, "Automatic Calibration of Piezoelectric Bed-Leaving Sensor Signals Using Genetic Network Programming Algorithms," Algorithms, vol. 14, no. 4, p. 117, 2021. [Online]. Available: https://www.mdpi.com/1999-4893/14/4/117

    2021

  21. [29]

    Ramezanian, A

    R. Ramezanian, A. Peymanfar, and S. B. Ebrahimi, "An integrated framework of genetic network programming and multi-layer perceptron neural network for prediction of daily stock return: An application in Tehran stock exchange market," Applied Soft Computing, vol. 82, p. 105551,...

    2019 doi

  22. [30]

    Variable -Size Genetic Network Programming for Portfolio Optimization with Trading Rules,

    F. Köhnke and C. Borgelt, "Variable -Size Genetic Network Programming for Portfolio Optimization with Trading Rules," in International Conference on the Applications of Evolutionary Computation (Part of EvoStar) , 2025: Springer, pp. 287-304

    2025

  23. [31]

    R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction (no. 1). A Bradford Book, 2018

    2018

  24. [32]

    Genetic Network Programming with Sarsa Learning Based Nonuniform Mutation,

    Q. Meng, S. Mabu, and K. Hirasawa, "Genetic Network Programming with Sarsa Learning Based Nonuniform Mutation," in IEEE International Conference on Systems, Man and Cybernetics, 10-13 Oct. 2010 2010, pp. 1273- 1278, doi: 10.1109/ICSMC.2010.5642421

    2010 doi

  25. [33]

    Robust Genetic Network Programming using SARSA Learning for autonomous robots,

    P. Sung Gil, S. Mabu, and K. Hirasawa, "Robust Genetic Network Programming using SARSA Learning for autonomous robots," in ICCAS-SICE, 18-21 Aug. 2009 2009, pp. 523-527

    2009

  26. [34]

    Genetic Network Programming with Reinforcement Learning Using Sarsa Algorithm,

    S. Mabu, H. Hatakeyama, K. Hirasawa, and H. Jinglu, "Genetic Network Programming with Reinforcement Learning Using Sarsa Algorithm," in IEEE International Conference on Evolutionary Computation, 0-0 0 2006, pp. 463-469, doi: 10.1109/CEC.2006.1688346

    2006 doi arXiv

  27. [35]

    Trading rules on stock markets using genetic network programming with sarsa learning,

    Y. Chen, S. Mabu, K. Hirasawa, and J. Hu, "Trading rules on stock markets using genetic network programming with sarsa learning," presented at the Proceedings of the 9th annual conference on Genetic and evolutionary computation, London, England, 2007

    2007

  28. [36]

    Genetic Network Programming with Reinforcement Learning and Its Application to Creating Stock Trading Rules,

    Y. Chen, S. Mabu, and K. Hirasawa, "Genetic Network Programming with Reinforcement Learning and Its Application to Creating Stock Trading Rules," in Machine Learning: InTech, 2009

    2009

  29. [37]

    A Learning Classifier System Based on Genetic Network Programming,

    X. Li and K. Hirasawa, "A Learning Classifier System Based on Genetic Network Programming," in IEEE International Conference on Systems, Man, and Cybernetics , 13 -16 Oct. 2013 2013, pp. 1323 -1328, doi: 10.1109/SMC.2013.229

    2013 doi

  30. [38]

    Niching genetic network programming with rule accumulation for decision making: An evolutionary rule -based approach,

    X. Li, M. Yang, and S. Wu, "Niching genetic network programming with rule accumulation for decision making: An evolutionary rule -based approach," Expert Systems with Applications, vol. 114, pp. 374 -387, 2018, doi: https://doi.org/10.1016/j.eswa.2018.07.041

    2018 doi

  31. [39]

    Genetic Network Programming with Estimation of Distribution Algorithms for class association rule mining in traffic prediction,

    X. Li, S. Mabu, H. Zhou, K. Shimada, and K. Hirasawa, "Genetic Network Programming with Estimation of Distribution Algorithms for class association rule mining in traffic prediction," in IEEE Congress on Evolutionary Computation, 18-23 July 2010 2010, pp. 1-8, doi: 10.1109/CEC...

    2010 doi

  32. [40]

    A Novel Graph-Based Estimation of the Distribution Algorithm and its Extension Using Reinforcement Learning,

    X. Li, S. Mabu, and K. Hirasawa, "A Novel Graph-Based Estimation of the Distribution Algorithm and its Extension Using Reinforcement Learning," IEEE Transactions on Evolutionary Computation, vol. 18, no. 1, pp. 98-113, 2014, doi: 10.1109/TEVC.2013.2238240

    2014 doi

  33. [41]

    A continuous estimation of distribution algorithm by evolving graph structures using reinforcement learning,

    X. Li, B. Li, S. Mabu, and K. Hirasawa, "A continuous estimation of distribution algorithm by evolving graph structures using reinforcement learning," in 2012 IEEE Congress on Evolutionary Computation, 10-15 June 2012 2012, pp. 1-8, doi: 10.1109/CEC.2012.6256481

    2012 doi

  34. [42]

    Adaptive Genetic Network Programming,

    X. Li, W. He, and K. Hirasawa, "Adaptive Genetic Network Programming," in IEEE Congress on Evolutionary Computation, 6-11 July 2014 2014, pp. 1808-1815, doi: 10.1109/CEC.2014.6900290

    2014 doi

  35. [43]

    Graph structure optimization of Genetic Network Programming with ant colony mechanism in deterministic and stochastic environments,

    M. Roshanzamir, M. Palhang, and A. Mirzaei, "Graph structure optimization of Genetic Network Programming with ant colony mechanism in deterministic and stochastic environments," Swarm and Evolutionary Computation, vol. 51, p. 100581, 2019, doi: https://doi.org/10.1016/j.swevo....

    2019 doi

  36. [44]

    Efficiency improvement of genetic network programming by tasks decomposition in different types of environments,

    M. Roshanzamir, M. Palhang, and A. Mirzaei, "Efficiency improvement of genetic network programming by tasks decomposition in different types of environments," Genetic Programming and Evolvable Machines, vol. 22, no. 2, pp. 229-266, 2021/06/01 2021, doi: 10.1007/s10710-021-09402-y

    2021 doi

  37. [45]

    Chapter Four - Situation-based genetic network programming to solve agent control problems,

    M. Roshanzamir and M. Roshanzamir, "Chapter Four - Situation-based genetic network programming to solve agent control problems," in Advances in Computers, vol. 135, A. Biswas, A. P. Tonda, R. Patgiri, and K. K. Mishra Eds.: Elsevier, 2024, pp. 77-97

    2024

  38. [46]

    Enhancing Multiagent Genetic Network Programming Performance Using Search Space Reduction,

    A. Kohan, M. Roshanzamir, and R. Alizadehsani, "Enhancing Multiagent Genetic Network Programming Performance Using Search Space Reduction," in International Conference on the Dynamics of Information Systems , 2025: Springer, pp. 94-105

    2025

  39. [47]

    Functionally distributed systems using parallel Genetic Network Programming,

    Y. Zhang, X. Li, Y. Yang, S. Mabu, Y. Jin, and K. Hirasawa, "Functionally distributed systems using parallel Genetic Network Programming," in Proceedings of SICE Annual Conference 2010, 18-21 Aug. 2010 2010, pp. 2626-2630

    2010

  40. [48]

    Graph-augmented large language model agents: Current progress and future prospects,

    Y. Liu, G. Zhang, K. Wang, S. Li, and S. Pan, "Graph-augmented large language model agents: Current progress and future prospects," arXiv preprint arXiv:2507.21407, 2025

    2025 arXiv

  41. [49]

    Graphs meet ai agents: Taxonomy, progress, and future opportunities,

    Y. Bei et al., "Graphs meet ai agents: Taxonomy, progress, and future opportunities," arXiv preprint arXiv:2506.18019, 2025

    2025 arXiv

  42. [50]

    AgentKit: structured LLM reasoning with dynamic graphs,

    Y. Wu et al., "AgentKit: structured LLM reasoning with dynamic graphs," arXiv preprint arXiv:2404.11483, 2024

    2024 arXiv

  43. [51]

    Scaling graph chain-of-thought reasoning: A multi-agent framework with efficient llm serving,

    C. Huan et al., "Scaling graph chain-of-thought reasoning: A multi-agent framework with efficient llm serving," arXiv preprint arXiv:2511.01633, 2025

    2025

  44. [52]

    Graph counselor: Adaptive graph exploration via multi -agent synergy to enhance llm reasoning,

    J. Gao et al., "Graph counselor: Adaptive graph exploration via multi -agent synergy to enhance llm reasoning," in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 24650-24668

    2025

  45. [53]

    Graphagent: Agentic graph language assistant,

    Y. Yang, J. Tang, L. Xia, X. Zou, Y. Liang, and C. Huang, "Graphagent: Agentic graph language assistant," in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , 2025, pp. 26360 - 26379

    2025

  46. [54]

    Can graph learning improve planning in llm-based agents?,

    X. Wu et al., "Can graph learning improve planning in llm-based agents?," Advances in Neural Information Processing Systems, vol. 37, pp. 5338-5383, 2024

    2024

  47. [55]

    Scalable and accurate graph reasoning with llm -based multi-agents,

    Y. Hu, R. Lei, X. Huang, Z. Wei, and Y. Liu, "Scalable and accurate graph reasoning with llm -based multi-agents," arXiv preprint arXiv:2410.05130, 2024

    2024

  48. [56]

    AgentRouter: A Knowledge -Graph-Guided LLM Router for Collaborative Multi -Agent Question Answering,

    Z. Zhang et al., "AgentRouter: A Knowledge -Graph-Guided LLM Router for Collaborative Multi -Agent Question Answering," arXiv preprint arXiv:2510.05445, 2025

    2025

  49. [57]

    Evoflow: Evolving diverse agentic workflows on the fly,

    G. Zhang et al., "Evoflow: Evolving diverse agentic workflows on the fly," arXiv preprint arXiv:2502.07373, 2025

    2025 arXiv

  50. [58]

    KnowThyself: An Agentic Assistant for LLM Interpretability,

    S. Prasai, M. Du, Y. Zhang, and F. Yang, "KnowThyself: An Agentic Assistant for LLM Interpretability," in Proceedings of the AAAI Conference on Artificial Intelligence, 2026, vol. 40, no. 48, pp. 41661-41663

    2026

  51. [59]

    Comparing some graph crossover in genetic network programming,

    H. Katagiri, K. Hirasawa, H. Jinglu, and J. Murata, "Comparing some graph crossover in genetic network programming," in Proceedings of the 41st SICE Annual Conference. SICE 2002. , 5-7 Aug. 2002 2002, vol. 2, pp. 1263-1268 vol.2, doi: 10.1109/SICE.2002.1195369

    2002 doi arXiv

  52. [60]

    Genetic Network Programming with Simplified Genetic Operators,

    X. Li, W. He, and K. Hirasawa, "Genetic Network Programming with Simplified Genetic Operators," in Neural Information Processing: 20th International Conference, ICONIP 2013, Daegu, Korea, November 3 -7, 2013. Proceedings, Part II , M. Lee, A. Hirose, Z. -G. Hou, and R. M. Kil ...

    2013

  53. [61]

    Revisiting genetic network programming (gnp): towards the simplified genetic operators,

    X. Li, H. Yang, and M. Yang, "Revisiting genetic network programming (gnp): towards the simplified genetic operators," IEEE Access, vol. 6, pp. 43274-43289, 2018

    2018

  54. [62]

    Children are more exploratory and learn more than adults in an approach -avoid task,

    E. G. Liquin and A. Gopnik, "Children are more exploratory and learn more than adults in an approach -avoid task," Cognition, vol. 218, p. 104940, 2022

    2022

  55. [63]

    Understanding explore-exploit dynamics in child development: current insights and future directions,

    S. Kim and S. M. Carlson, "Understanding explore-exploit dynamics in child development: current insights and future directions," Frontiers in Developmental Psychology, vol. 2, p. 1467880, 2024

    2024

  56. [64]

    Early adversity and the development of explore –exploit tradeoffs,

    W. E. Frankenhuis and A. Gopnik, "Early adversity and the development of explore –exploit tradeoffs," Trends in Cognitive Sciences, vol. 27, no. 7, pp. 616-630, 2023

    2023

  57. [65]

    Introducing the Tileworld: experimentally evaluating agent architectures,

    M. Pollack and M. Ringuette, "Introducing the Tileworld: experimentally evaluating agent architectures," environment, pp. 183-189, 1990

    1990

  58. [66]

    D. D. Wackerly, W. Mendenhall, and R. L. Scheaffer, Mathematical statistics with applications. Thomson Brooks/Cole Belmont, CA, 2008

    2008

  59. [67]

    Reasoning with exploration: An entropy perspective,

    D. Cheng et al., "Reasoning with exploration: An entropy perspective," in Proceedings of the AAAI Conference on Artificial Intelligence, 2026, vol. 40, no. 36, pp. 30377-30385

    2026

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.