IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 1, February 2026, pp. 536 ∼ 546 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i1.pp536-546 ❒ 536 AI-po wer ed hub optimization: a r einf or cement lear ning and graph-based appr oach to scalable blockchain netw orks Kassem Danach 1 , Hassan Rk ein 1 , Alaaeddine Ramadan 2 , Hassan Harb 3 , Bassam Hamdar 1 1 Basic and Applied Sciences Research Center , Al Maaref Uni v ersity , Beirut, Lebanon 2 Colle ge of Engineering and Computing, American Uni v ersity of Bahrain, Rif f a, Bahrain 3 Colle ge of Engineering and T echnology , American Uni v ersity of the Middle East, Eg aila, K uw ait Article Inf o Article history: Recei v ed Jun 9, 2025 Re vised No v 19, 2025 Accepted Dec 15, 2025 K eyw ords: Articial intelligence Blockchain Combinatorial optimization Graph neural netw orks Hub location Reinforcement learning ABSTRA CT Blockchain netw orks f ace persistent scalability chal lenges, including netw ork congestion, high latenc y , and transaction costs. T o address these limitations, this study proposes an AI-dri v en hub location optimization frame w ork that inte grates reinforcement learning (RL), mix ed inte ger linear programming (MILP), and graph neural netw orks (GNNs). The RL-based hub selection dynamically identies optimal supernode placement, while MILP ensures cost-ef cient transaction routing, and GNNs predict o w patterns for proacti v e congestion management. Experimental results on Ethereum and Bitcoin datasets demonstrate s ignicant impro v ements, including a 58.6% reduction in transaction latenc y , 28.9% g as fee sa vings, and 41.5% congestion reduction. Be yond performance g ains, statistical t ests conrm the signicance of these impro v ements, and ablation studies highlight the complementary role of each component. This is an open access article under the CC BY -SA license . Corresponding A uthor: Alaaeddine Ramadan Colle ge of Engineering and Computing, American Uni v ersity of Bahrain Rif f a, Bahrain Email: alaaeddine.ramadan@aubh.edu.bh 1. INTR ODUCTION Blockchain enables decentralized, secure, and transparent transactions across multiple sectors, including nance, supply chain management, and decentralized autonomous or g anization (D A Os). Y et, scalability is sues persist due to the hea vy computational demands of consensus mechanisms such as proof-of-w ork (PoW) and proof-of-stak e (PoS) [1]. W ith Bitcoin and Ethereum processing only about 7 and 30 TPS compared to V isa’ s 65,000 TPS [2], de v eloping more ef cient blockchain management solutions is essential. Current scalability solutions include Layer -2 frame w orks lik e the lightning netw ork and plasma, which of oad transactions to secondary layers b ut introduce security risks and require si gnicant architectural changes [3]. Another approach emplo ys hub-based netw ork management, where selected high-capacity nodes impro v e throughput and reduce congestion [4]. Ho we v er , peer -to-peer broadcasting remains inef cient since all nodes redundantly v alidate and store transactions, while g as fee–based prioritization and the absence of adapti v e routing and load-balancing mechanisms further hinder performance. These challenges highlight the necessity for AI-dri v en optimization to enable intelligent, real-time blockchain management. Inte grating machine learning (ML), reinforcement learning (RL), and combinatorial opt imization presents a promising solution to blockchain scalability challenges. Graph neural netw orks (GNNs) can analyze transaction graphs to predict congestion, while deep reinforcement learning (DRL) optimizes hub J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 537 placement based o n dynamic traf c patterns. Coupled with mix ed inte ger linear programming (MILP) for precise routing optimization, this combination enables predicti v e, adapti v e, and ef cient netw ork management, reducing latenc y , enhancing scalability , and balancing computational loads. This study introduces an AI-dri v en hub location optimiza tion frame w ork combining RL, MILP , and GNNs to enhance blockchain scalability and ef cienc y . It utilizes deep Q-netw orks (DQN) and proxi mal polic y optimization (PPO) for adapti v e hub allocation, a MILP model to minimize g as fees and optimize routing, and a temporal graph con v olutional netw orks (T -GCN) to predict congestion patterns. Compared with peer -to-peer and x ed hub approaches, the proposed method achie v es a 58.6% latenc y reduction, 28.9% lo wer g as fees, and 41.5% better congestion control, with statistically signicant results ( p < 0 . 01 ). Ablation and feature analyses conrm the complementary roles and interpret ability of the frame w ork’ s components, of fering a scalable and intelligent foundation for decentralized systems such as decentralized nance (DeFi), supply chains, and smart contracts. 2. RELA TED W ORK Blockchain research increasingly focus es on scalability , ef cienc y , and netw ork management. Hybrid approaches combining RL, optimization, and graph models sho wing promise b ut remaining partiall y inte grated. Blockchain also enhances transparenc y , auditability , and re gulatory compliance in cryptocurrenc y accounting, supporting more reliable nancial reporting. 2.1. Blockchain netw ork scalability and transaction pr ocessing strategies The PoW consensus mechanism used in Bitcoin ensures security and decentralization b ut introduces high computational costs and delays in transaction conrmation as well as scalability issues [5]. Se v eral approaches ha v e been proposed to address such challenges. Layer -2 scaling solutions, such as the lightning netw ork for Bitcoi n and plasma for Ethereum, enable of f-chain transactions to alle viate congesti on on the main chain [6], [7]. Additionally , shardi ng techniques ha v e been e xplored to partition blockchain netw orks into smaller , more manageable units that process transactions in parallel [8]. While these methods impro v e scalability , the y introduce ne w challenges, inc luding security vulnerabilities and increased netw ork comple xity . An alternati v e approach in v olv es intelligent transaction r o ut ing and netw ork optimization, where k e y nodes (hubs) play a central role in managing transaction o w and reducing congestion [9]. The hub location problem (HLP) is a well-established combinatorial optimization challenge that seeks to determine the optimal placement of hubs in a netw ork to minimize cost and maximize ef cienc y [10]. Recent research suggests that dynamic hub selection based on netw ork conditions and transaction o w can further enhance ef cienc y [11]. Ho we v er , e xisting studies ha v e primarily focused on static hub placement, which f ails to account for real-time uctuations in netw ork demand. 2.2. Machine lear ning techniques f or blockchain management ML has been increasingly applied to blockchain systems for tasks such as fraud detection, s mart contract security , and netw ork optimization [12], [13]. Ho we v er , ML models often require lar ge labeled datasets and do not adapt well to the dynamic nature of blockchain transactions. Deep learning techniques ha v e sho wn promise in forecasting blockchain transaction trends [14]. These models can be used to predict netw ork congestion and transaction demand, allo wing for more ef cient transaction routing and resource allocation. Despite these adv ancements, traditional deep learning methods do not inherently capture the graph-based structure of blockchain transactions. As a result, GNNs ha v e emer ged as a po werful tool for analyzing blockchain transaction data. GNNs can model transaction relationships, detect anomalies, and predict netw ork congestion hotspots [15]. In this research, a GNN-based transaction prediction model is inte grated into the proposed hub location optimization frame w ork to enhance blockchain scalability and ef cienc y . 2.3. Reinf or cement lear ning and graph neural netw orks in decentralized systems RL has been widely used in netw ork optimization problems, including dynamic resource alloc ation, traf c management, and decentralized control systems [16], [17]. RL-based techniques of fer the adv antage of learning optimal netw ork management policies through continuous interaction with the blockchain en vironment. Unlik e rule-based methods, RL models can adapt to changing transaction patterns and dynamically adjus t netw ork parame ters in rea l-time [18]. Recent research has also e xplored the combi nation of GNNs and RL for decentralized decision-making [19]. GNNs pro vide spatial insights into transaction beha vior , while RL enables the system t o mak e autonomous decisions for hub selection and transaction routing. By AI-power ed hub optimization: a r einfor cement learning and gr aph-based appr oac h to ... (Kassem Danac h) Evaluation Warning : The document was created with Spire.PDF for Python.
538 ❒ ISSN: 2252-8938 inte grating these techniques, thi s study aims to de v elop an adapti v e hub location optimization frame w ork that balances transaction load, reduces congestion, and enhances blockchain ef cienc y . As summarized in T able 1, prior h ybrid approaches ha v e made v aluable contrib utions by combining AI and optimization techniques. Ho we v er , the y remain limited to partial inte grations (e.g., RL+MILP or GNN+heuristics ). Our w ork adv ances this line of research by unifying RL for adapti v e hub selection, MILP for cost-optimal routing, and GNNs for congestion prediction, thereby addressing blockchain scalability in a holistic manner . T able 1. Comparati v e benchmarking of h ybrid AI–optimization models for netw ork and hub optimization Ref. Domain T echniques Data / scale Reported impro v ements K e y limitations [20] Logistics hub location RL + MILP 500–1000 node s Cost reduction ≈ 20% Static hub placement, no dynamic adaptation [21] T raf c routing GNN + Heuristics Road netw ork simulations Congestion reduction ≈ 30% No cost-optimal routing, limited scalability [22] Communication netw orks RL + GNN Simulated netw ork topology Latenc y reduction ≈ 25% Lacks formal optimization (MILP) This w ork Blockchain netw orks RL (DQN/PPO) + MILP + GNN Ethereum & Bitcoin datasets Latenc y ↓ 58.6%, Gas fees ↓ 28.9%, Congestion ↓ 41.5% Higher computational o v erhead for RL/MILP inte gration 3. PR OPOSED METHODOLOGY This research introduces an AI-dri v en hub location optimization frame w ork. The frame w ork inte grates RL, MILP , and GNNs to enhance blockchain transaction ef cienc y and scalability . The interaction between these components is summarized in Figure 1, which illustrates the w orko w of the proposed system. Figure 1. W orko w of the AI-dri v en hub optimization frame w ork 3.1. Reinf or cement lear ning-based hub selection The hub selection problem in blockchain netw orks is formulated as a Mark o v decision process (MDP), where an RL agent optimizes hub node placement based on real-time transaction traf c and netw ork conditions. The RL model is designed to optimize hub selection by representing the blockchain netw ork as a MDP . This formulation ensures that the agent can percei v e the en vironment in terms of state, action, and re w ard, which are the core components of RL (Algorithm 1). The state space ( S ) includes: transaction v olume per node, node processing po wer , block conrmation times, transaction fees, and congestion forecasts from the GNN. The action space ( A ) consists of selecting one or more nodes as hubs. The re w ard function ( R ) assigns positi v e re w ards for higher throughput, reduced latenc y , and lo wer g as fees, while applying penalties for congestion or imbalance. The RL agent is trained for 5,000 episodes u s ing an ϵ -greedy e xploration strate gy with ϵ decaying Int J Artif Intell, V ol. 15, No. 1, February 2026: 536–546 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 539 from 1.0 to 0.1. T ar get netw orks are updated e v ery 100 steps, with a replay b uf fer of 50,000. The Adam optimizer is applied with learning rate 10 − 4 and bat ch size 64. PPO is trained wit h tw o hidden layers (128, 64), clip ratio 0.2, and entrop y re gularization. Con v er gence is determined when the mo ving a v erage of cumulati v e re w ards stabilizes o v er 50 episodes. Algorithm 1 RL training loop for hub selection 1: Initialize replay b uf fer D , netw orks Q , Q ′ 2: Set α = 10 − 4 , batch size = 64 , ϵ = 1 . 0 → 0 . 1 3: f or episode = 1 to 5000 do 4: Initialize state s 0 (transactions, node load, GNN predictions) 5: f or each step t do 6: Choose action a t using ϵ -greedy strate gy 7: Ex ecute hub allocation a t , observ e re w ard r t , ne xt state s t +1 8: Store ( s t , a t , r t , s t +1 ) in b uf fer D 9: Update Q with Adam optimizer; e v ery 100 steps update Q ′ 10: end f or 11: Decay ϵ gradually , check con v er gence criteria 12: end f or 3.2. Mixed integer linear pr ogramming f or transaction r outing Once hubs are selected, transaction routing optimization is performed using MILP . The objecti v e is to minimize routing costs (latenc y , g as fees, o v erhead) subject to hub capacity and assignment constraints. min X i ∈ N X j ∈ H C ij X ij (1) Accordingly , t w o constraints are dened: (1) each node is assigned to one hub ( P j ∈ H X ij = 1 ), and (2) hub capacity ( P i ∈ N X ij ≤ C j ). Then, we relax hub capacity constraints with multipliers λ j , decompose the MILP into parallel subproblems, and update multipliers using subgradient optimization (Algorithm 2). This reduces runtime while maintaining near -optimal routing. The trade-of f is that strict optimality may not be guaranteed, b ut ef cienc y g ains are critical for blockchain-scale netw orks. Algorithm 2 MILP with Lagrangian relaxation 1: Initialize multipliers λ j ≥ 0 , step size η 2: r epeat 3: Decompose MILP into routing subproblems 4: Solv e each subproblem in parallel for X ij 5: Compute violations v j = P i X ij − C j 6: Update multipliers: λ j ← max(0 , λ j + η v j ) 7: until con v er gence 8: Return near -optimal X ij with multipliers λ j 3.3. Graph neural netw orks f or transaction o w pr ediction W e model the blockchain as a temporal graph G = ( V , E , T ) , where nodes are addresses, edges are transactions, and T encodes block timestamps. Features include transaction v olume, fees, node de gree, and temporal clustering. The architecture of the T -GCN consists of tw o graph layers (64 units each) + long term short memory (LSTM) (128 units), dropout 0.3. T rained for 200 epochs with Adam ( 5 × 10 − 4 ), batch size 64, using mean squared error (MSE) loss and early stopping. AI-power ed hub optimization: a r einfor cement learning and gr aph-based appr oac h to ... (Kassem Danac h) Evaluation Warning : The document was created with Spire.PDF for Python.
540 ❒ ISSN: 2252-8938 4. EXPERIMENT AL SETUP AND D A T ASET 4.1. Blockchain datasets T o e v aluate the proposed AI-dri v en hub location optimization frame w ork, we utilize publicly a v ailable blockchain transaction datasets from the Ethereum and Bitcoin netw orks: i) Ethereum dataset: we utilize the Ethereum transaction dataset from the Etherscan API and the BigQuery Ethereum dataset [23], [24]. These datasets contain detailed Ethereum blockchain transactions, including: – T ransaction hashes, sender and recei v er addresses – Gas fees, g as limits, and base fees – Smart contract interactions and ERC-20 tok en transfers – Block timestamps and miner details ii) Bitcoin dataset: the Bitcoin transaction data is obtained from the Bitcoin blockchain data repository and the Kaggle Bitcoin dataset [25], [26]. These datasets pro vide: – Bitcoin transaction records, including sender and recei v er addresses – T ransaction sizes, input/output v alues, and mining fees – Block conrmation times and mempool w aiting periods – Unspent transaction output (UTXO) analysis for scalability assessment 4.2. F eatur e engineering The follo wing k e y features are e xtracted to capture essential transaction patterns and optim ize blockchain transaction ef cienc y: – T ransaction v olume per block ( T bl ock ) : measures the number of transactions processed within a single bl o c k. This feature directly relates to throughput and is critical for assessing scalability . – A v erage transaction conrmation time ( T conf ir m ) : calculates the mean tim e tak en for transactions to be v alidated and recorded. It reects latenc y , a k e y performance indicator for user e xperience. – Gas price and transaction fee uctuations ( F g as ) : tracks the v ariability in transaction costs o v er time. High v olatility in g as fees pro vides signals for congestion forecasting. – De gree c entrality of transaction nodes ( D centr al ity ) : determines the inuence of nodes based on the number of transactions the y process. Central nodes are more lik ely to become hubs. – T emporal transaction clustering patterns ( C tempor al ) : identies repeating transaction beha viors o v er s pecic interv als. These patterns capture diurnal and c yclical demand shifts impacting netw ork load. 4.3. Simulation en vir onment and hard war e specications The AI-dri v en blockchain management s y s tem is e v aluated based on the follo wing softw are and hardw are specications: – PyT orch: used for deep learning model de v elopment. – Netw orkX: f acilitates graph-based transaction netw ork processing. – Gurobi: handles MILP-based optimization for transaction routing. – CPU: Intel Xeon Silv er 4214 (2.2 GHz, 12 cores) – GPU: NVIDIA R TX 3090 (24 GB VRAM) – RAM: 128 GB DDR4 – Storage: 4 TB SSD Additionally , the Ethereum blockchain emulator (Ganache) is used to simulate transaction e x ecution, measure real-w orld transaction costs, and v alidate the ef fecti v eness of the proposed hub optimization frame w ork. This emulator pro vides a controlled en vironment to test the optimization strate gies before v alidating them ag ainst real blockchain traces. 5. RESUL TS AND PERFORMANCE EV ALU A TION 5.1. Ev aluation metrics T o assess the ef fecti v eness of the proposed AI-dri v en hub location optimizati on frame w ork, we ut ilize the follo wing k e y performance metrics: Int J Artif Intell, V ol. 15, No. 1, February 2026: 536–546 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 541 – T ransaction latenc y (TL): measures the a v erage time tak en for transactions to be conrmed, dened as: T L = 1 N N X i =1 ( T conf ir m,i − T submit,i ) (2) where T conf ir m is the timestamp of block conrmation, and T submit is the transaction submission time. – Gas fee optimization (GFO): quanties the reduction in g as fees achie v ed by AI-dri v en hub selection compared to traditional blockchain routing: GF O = F basel ine − F optimiz ed F basel ine × 100% (3) where F basel ine is the a v erage g as fee before optimization, and F optimiz ed is the g as fee after optimization. – Netw ork congestion reduction (NCR): e v aluates the ef cienc y of the hub selection process in reducing netw ork congestion, dened as: N C R = 1 | H | X j ∈ H 1 − T cong ested,j T total ,j × 100% (4) where T cong ested,j represents the number of delayed transactions at hub j , and T total ,j is the total transact ions processed by hub j . 5.2. Benchmarking against additional baselines T o pro vide a stronger comparati v e perspecti v e, we e v aluated the proposed RL–MILP–GNN frame w ork ag ainst additional baselines including static hub placement, greedy routing, and traditional blockchain scaling methods (e.g., layer -2 solutions and sharding). T able 2 presents the results. The proposed frame w ork consistently outperforms these methods, reducing latenc y by 36.5% relati v e to static placement and impro ving congestion control by 29.4% compared to greedy routing. T able 2. Performance benchmarking ag ainst baseline strate gies Model Latenc y (sec.) GFO (%) Congestion reduction (%) Static hub placement 8.3 15.2 22.1 Greedy routing 7.5 18.4 28.7 T raditional layer -2 / sharding 6.9 20.6 31.2 Proposed RL–MILP–GNN 5.3 28.9 41.5 5.3. Sensiti vity analysis and r ob ustness testing T o e v aluate the rob ustness of the AI-dri v en hub selection model, we tested its performance under v arying transaction loads. Figure 2 illustrates transaction latenc y trends across dif ferent netw ork conditions. K e y ndings from the rob ustness analysis include: – The RL-based hub selection dynamically adjusts to netw ork congestion, ensuring minimal transaction delay . – Performance remains stable e v en under 200% increased transaction v olume, conrming the model’ s scalability . – The MILP-based transaction routing ef fecti v ely distrib utes w orkload across hubs, pre v enting netw ork bottlenecks. 5.4. Hub w orkload distrib ution T o assess w orkload balancing ef cienc y , we comput ed the v ariance in w orkload across hubs. T able 3 presents w orkload v ariance under dif ferent transaction loads. Figure 3 visualizes w orkload distrib ution across scenarios. These results conrm that the frame w ork maintains balanced w orkload distrib ution in most conditions, with slight v ariance increases under e xtreme transaction loads. AI-power ed hub optimization: a r einfor cement learning and gr aph-based appr oac h to ... (Kassem Danac h) Evaluation Warning : The document was created with Spire.PDF for Python.
542 ❒ ISSN: 2252-8938 Figure 2. Sensiti vity analysis of AI-dri v en hub selection under v arying transaction loads T able 3. W orkload v ariance across hubs T ransaction v olume scenario A v erage w orkload v ariance Lo w v olume (50%) 2.15 Standard v olume (Baseline) 1.47 High v olume (200%) 3.02 Figure 3. W orkload distrib ution across hubs for dif ferent transaction v olumes 5.5. Scalability analysis T o assess system scalability , we tested the frame w ork under increasing transaction loads (50%, 100%, and 200%). T able 4 presents the performance results. While the frame w ork ef ciently handles standard loads, performance de grades under e xtreme transaction v olumes (200%), highlighting opportunities for adapti v e resource scaling. These trends are visualized in Figure 4, which illustrates the relationship between transaction load, latenc y , and throughput. Notably , latenc y increases sharply at 200% load, while throughput drops signicantly , conrming the need for dynamic resource optimization under high demand. Int J Artif Intell, V ol. 15, No. 1, February 2026: 536–546 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 543 T able 4. Scalability test results T ransaction load latenc y (sec.) Throughput (TPS) Resource utilization (%) Lo w (50%) 4.2 1200 45.3 Standard (100%) 6.8 950 68.1 High (200%) 12.5 720 91.4 Figure 4. Scalability test results sho wing transaction latenc y and throughput 5.6. Ener gy efciency analysis The AI-dri v en hub optimization frame w ork achie v ed a substantial reduction in ener gy consum p t ion, demonstrating its ef cienc y be yond just performance metrics. This ener gy ef cienc y w as quantied using the fundamental equation E = P × T , where E is ener gy (kWh), P is the measured po wer dra w (W), and T is the total transaction processing time (s). The results, presented in T able 5, conrm that the optimized system lo wered ener gy usage to 3.2 kWh, a 40.7% decrease compared to the 5.4 kWh consumed by the traditional baseline. T able 5. Ener gy consumption comparison Scenario Po wer (W) Ener gy consumption (kWh) Baseline (traditional) 1800 5.4 AI-optimized hub selection 1200 3.2 5.7. Statistical signicance testing T o v alidate the rob ustness of the proposed frame w ork, we applied non-parametric statistical tes ts across all baseline and comparati v e models. A Friedman rank test w as rst conducted to assess o v erall dif ferences, follo wed by pairwise W ilcoxon signed-rank tests with Bonferroni correction. The results conrm that the proposed AI-dri v en frame w ork achie v es statistically signicant impro v ements ( p < 0 . 01 ) in transaction latenc y , g as fee optimization, and congestion reduction compared to both peer -to-peer baselines and x ed hub selection models. This pro vides strong e vidence that the reported impro v ements are not due to random v ariation. 5.8. F eatur e impact analysis Finally , we analyzed the impact of input features on model predictions using Shaple y additi v e e xplanations (SHAP) v alues. The results sho w that transaction v olume per block ( T bl ock ) and g as fee uctuations ( F g as ) are the most inuential predictors of hub selection and routing performance. De gree centrality ( D centr al ity ) and temporal clustering patterns ( C tempor al ) also contrib ute, b ut with lo wer relati v e AI-power ed hub optimization: a r einfor cement learning and gr aph-based appr oac h to ... (Kassem Danac h) Evaluation Warning : The document was created with Spire.PDF for Python.
544 ❒ ISSN: 2252-8938 importance. This feat ure impact analysis pro vides interpretable insights into the dri v ers of scalability in blockchain netw orks, strengthening the transparenc y and interpretability of the proposed AI-dri v en frame w ork. 6. CONCLUSION AND FUTURE W ORK This study introduces a no v el AI-dri v en frame w ork that unies RL, MILP , and GNNs to si gnicantly enhance blockchain scalability , achie ving notable reductions in transaction latenc y , g as fees, a nd netw ork congestion through dynamic hub selection, optimized routing, and predicti v e congestion management. Ho we v er , the frame w ork f aces limitations, including high computational demands, partial centralization risks, and assumptions about netw ork stability and hub capacity . Future w ork will therefore focus on enhancing adaptability through meta-learning, inte grating with layer -2 solutions, and v alidating the approach across di v erse blockchain platforms and real-w orld applications lik e DeFi and D A Os to adv ance the de v elopment of more intelligent and scalable decentralized systems. FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Kassem Danach ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Hassan Rk ein ✓ ✓ ✓ ✓ ✓ ✓ Alaaeddine Ramadan ✓ ✓ ✓ ✓ ✓ ✓ Hassan Harb ✓ ✓ ✓ ✓ ✓ ✓ ✓ Bassam Hamdar ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST Authors state no conict of interest. D A T A A V AILABILITY The data that support the ndings of this study are a v ailable from the rst author or the corresponding author , [AR], upon reasonable request. REFERENCES [1] S. Punitha and K. S. Preetha, “ A no v el inte gration of web 3.0 with h ybrid chaotic-hippo-optimized blockchain frame w ork for healthcare 4.0, ” Results in Engineering , v ol. 24, Dec. 2024, doi: 10.1016/j.rineng.2024.103528. [2] O. Ascue, O. V alle, and J. Santisteban, “BLOCKSA GE: Blockchain-based cloud architecture for sensiti v e data management in SMEs, ” Sustainability , v ol. 17, no. 4, Feb . 2025, doi: 10.3390/su17041352. [3] S. G. Sa v adatti, S. Krishnamoort h y , a nd R. Delhibab u, “Surv e y of distrib uted ledger technology (DL T) for secure and scalable computing, ” IEEE Access , v ol. 13, pp. 8393–8415, 2025, doi: 10.1109/A CCESS.2025.3528211. [4] I. S. Rao, M. L. M. Kiah, M. M. Hameed, and Z. A. Memon, “Scal ability of blockchain: a comprehensi v e re vie w and future research direction, ” Cluster Computing , v ol. 27, no. 5, pp. 5547–5570, Aug. 2024, doi: 10.1007/s10586-023-04257-7. [5] Z. U. Abadin and M. Syed, “ A pattern for proof of w ork consensus algorithm in blockchain, ” in 26th Eur opean Confer ence on P attern Langua g es of Pr o gr ams , Graz, Austria: A CM, July 2021, pp. 1–6, doi: 10.1145/3489449.3489994. Int J Artif Intell, V ol. 15, No. 1, February 2026: 536–546 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 545 [6] G. A. F . Rebello et al. , “ A surv e y on blockchain scalability: from hardw are t o layer -tw o protocols, ” IEEE Communications Surve ys & T utorials , v ol. 26, no. 4, pp. 2411–2458, 2024, doi: 10.1109/COMST .2024.3376252. [7] A. Gangw al, H. R. Gang a v alli, and A. Thirupathi, “ A surv e y of layer -tw o blockchain protocols, ” J ournal of Network and Computer Applications , v ol. 209, Jan. 2023, doi: 10.1016/j.jnca.2022.103539. [8] M. Zamani, M. Mo v ahedi, and M. Rayk o v a, “RapidChain: scaling blockchain via full sharding, ” in Pr oceedings of the 2018 A CM SIGSA C Confer ence on Computer and Communications Security , T oronto, Canada: A CM, Oct. 2018, pp. 931–948, doi: 10.1145/3243734.3243853. [9] A. A. Mazlan, S. M. Daud, S. M. Sam, H. Abas, S . Z. A. Rasid, and M. F . Y usof, “Scalability challenges in healthcare blockchain system—a systematic re vie w , ” IEEE Access , v ol. 8, pp. 23663–23673, 2020, doi: 10.1109/A CCESS.2020.2969230. [10] S. A. Alumur , J. F . Campbell, I. Contreras, B. Y . Kara, V . Mariano v , and M. E. O’K elly , “Perspecti v es on modeling hub location problems, ” Eur opean J ournal of Oper ational Resear c h , v ol. 291, no. 1, pp. 1–17, May 2021, doi: 10.1016/j.ejor .2020.09.039. [11] K. Danach, A. Rammal , I. Moukadem, H. Harb, and A. Nasser , “ Adv anced optimization in e-commerce logistics: combining m atheuristics with random forests for hub location ef cienc y , ” IEEE Access , v ol. 13, pp. 55915–55926, 2025, doi: 10.1109/A CCESS.2025.3550560. [12] F . Chen, H. W an, H. Cai, and G. Cheng, “Machine learning in/for blockchain: future and challenges, ” Canadian J ournal of Statistics , v ol. 49, no. 4, pp. 1364–1382, 2021, doi: 10.1002/cjs.11623. [13] K. Danach, H. Harb, A. Ramadan, and S. Haddad, “Enhancing multi-criteria decision-making in blockchain security: a h ybrid machine learning and PR OMETHEE approach, ” Engineering Resear c h Expr ess , v ol. 7, no. 3, pp. 1-22, Sep. 2025, doi: 10.1088/2631-8695/ae05eb . [14] J. W eng, J. W eng, J . Zhang, M. Li, Y . Zhang, and W . Luo, “DeepChain: auditable and pri v ac y-preserving deep le arning with blockchain-based incenti v e, ” IEEE T r ansactions on Dependable and Secur e Computing , v ol. 18, no. 5, pp. 2438-2455, 2021, doi: 10.1109/TDSC.2019.2952332. [15] X. Lin, J. W u, A. K. Bashir , J. Li, W . Y ang, and M. J. Piran, “Blockchain-based incenti v e ener gy-kno wledge trading in IoT : joint po wer transfer and AI design, ” IEEE Internet of Things J ournal , v ol. 9, no. 16, pp. 14685–14698, 2022, doi: 10.1109/JIO T .2020.3024246. [16] V . Mnih et al ., “Human-le v el control through deep reinforcement learning, ” Natur e , v ol. 518, no. 7540, pp. 529–533, Feb . 2015, doi: 10.1038/nature14236. [17] B. Han et al ., “Dynamic incenti v e design for federated learning based on consortium blockchain using a stack elber g g ame, ” IEEE Access , v ol. 12, pp. 160267–160283, 2024, doi: 10.1109/A CCESS.2024.3487585. [18] A. S. Almasoud, “Blockchain-based secure storage and sharing of medical data using machine learning, ” in 2023 T enth International Confer ence on Social Networks Analysis, Mana g ement and Security (SN AMS) , Ab u Dhabi, United Arab Emirates: IEEE, No v . 2023, pp. 1–4, doi: 10.1109/SN AMS60348.2023.10375435. [19] M. Seo, J. Kim, M. Y ou, S. Shin, and J. Kim, “gShock: a GNN-based ngerprinting system for permissioned blockchain netw orks o v er encrypted channels, ” IEEE Access , v ol. 12, pp. 146328–146342, 2024, doi: 10.1109/A CCESS.2024.3469583. [20] T .-H. Lee and M.-S. Kim, “RL-MILP solv er: a reinforcement learning approach for solving mix ed-inte ger linear programs with graph neural netw orks, ” 5th Annual AAAI W orkshop on AI to Acceler ate Science and Engineering (AI2ASE) , pp. 1–10, 2023. [21] T . Liu and H. Meidani, “End-to-end heterogeneous graph neural netw orks for traf c assi gnment, ” T r ansportation Resear c h P art C: Emer ging T ec hnolo gies , v ol. 165, 2024, doi: 10.1016/j.trc.2024.104695. [22] P . Almasan, J. S.-V arela, K. Rusek, P . B.-Ros, and A. C.-Apari cio, “Deep rei nforcement learning meets graph neural netw orks: e xploring a routing optimization use case, ” Computer Communications , v ol. 196, pp. 184–194, 2022, doi: 10.1016/j.comcom.2022.09.029. [23] N. S. K umar , S. Mirdula, P . Singh, T . Gayathri, J. S, and A. M, “ A paradigm shift in ethereum netw ork analysis through Google BigQuery ,portation, and visualization, ” 2024 International Confer ence on Emer ging Smart Computing and Informatics (ESCI) , pp. 1–7, 2024, doi: 10.1109/ESCI59607.2024.10497269. [24] A. Day and E. Medv ede v , “Ethereum i n BigQuery: a public dataset for smart contract analytics, ” Goo gle Cloud Blo g , Accessed: May 27, 2023. [Online]. A v ailable: https://cloud.google.com/blog/products/data-a nalytics/ethereum-bigquery-public-dataset-smart- contract-analytics [25] A. Das, G. Uddin, and G. Ruhe, “ An empirical study of blockchai n repos itories in GitHub, ” in The International Confer ence on Evaluation and Assessment in Softwar e Engineering 2022 , Gothenb ur g, Sweden: A CM, June 2022, pp. 211–220, doi: 10.1145/3530019.3530041. [26] Google BigQuery , M. Risdal, S. Dane, and A. Day , “Bitcoin blockchain historical data, ” Ka g gle , 2020, Accessed: May 27, 2023. [Online]. A v ailable: https://www .kaggle.com/datasets/bigquery/bitcoin-blockchain BIOGRAPHIES OF A UTHORS Kassem Danach recei v ed the Doctor of Philosoph y (Ph.D.) de gree in Computer Engineering from Ecole Centrale de Lille, France, in 2016. He has applied his e xpertise in data analyt ics, articial intelligence, machine learning, deep learning, educational technology , and b usiness analytics to v arious military and professional roles. He has made subst antial contrib utions to the eld of operational research and articial intelligence, earning him prestigious a w ards, including the best paper a w ard at LCIS in 2017 and AISD in 2019, and acti v e participation as a team member in the EU-funded SOCORR O project, from 2019 to 2021. He can be contacted at email: kassem.danach@mu.edu.lb . AI-power ed hub optimization: a r einfor cement learning and gr aph-based appr oac h to ... (Kassem Danac h) Evaluation Warning : The document was created with Spire.PDF for Python.