Özkan Canay
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Özkan Canay has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Cem Özkurt
Sakarya Uygulamali Bilimler Universitesi
3 shared publications - Halil Arslan
Sivas Cumhuriyet Üniversitesi
3 shared publications Ümit Kocabıçak
3 shared publications- Ahmet Dursun
Sakarya Uygulamali Bilimler Universitesi
1 shared publication - Beyza Sıla Velioğlu
Sakarya University of Applied Sciences
1 shared publication - Manoj Kumar Misra
Allenhouse Institute of Technology
1 shared publication
Latest Publications
- 28 September 2026
- Published by Dicle Universitesi Muhendislik Fakultesi Muhendislik Dergisi in Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi
- Vol. 17 (3)
A media-type-aware asynchronous framework for restoring heterogeneous stream endpoint collections
User-maintained Internet radio and IPTV-like playlists often contain unavailable, duplicate, redirected, malformed, or non-media endpoints. This article evaluates a reproducible, provenance-preserving Python framework for restoring heterogeneous stream endpoint collections through playlist-level media-type-aware validation, URL transformation, invalid-entry removal, and verified alternative selection. The framework parses TXT, M3U/M3U8, pipe-delimited, and AIMPPL4 inputs and normalizes station–URL pairs before applying asynchronous HTTP HEAD probing with a GET fallback whose response-body read is capped at one byte after HEAD 405/501. It then revalidates transformed URLs, selects one validated alternative per station, and exports machine-readable evidence. Across three repetitions on a fixed 400-endpoint sample in the primary cloud environment, a concurrency limit of 50 reduced mean validation time from 211.6 to 6.3 seconds, a 33.6-fold speedup. An independent repetition in a second environment showed the same timing direction. A separate full-corpus evaluation processed all 1,858 unique pairs in 31.14 seconds, validated 485 endpoints, and selected endpoints for 372 stations. In a four-configuration descriptive acceptance/repair analysis, the status-only/no-repair setting accepted 25.6 non-media 2xx responses on average, whereas the media-type/repair setting accepted 115.8 media-compliant endpoints and rescued 29.3. Within the sampled decoder strata, all 19 unique rescued URLs passed a short independent media probe and none of the 18 unique status-only non-media URLs did, although three of 20 validator-rejected URLs also passed. These results characterize time-bounded restoration of endpoint collections under the specified experimental conditions and do not establish recovery of remote services, sustained playback, or quality of experience.
- 28 September 2026
- Published by Dicle Universitesi Muhendislik Fakultesi Muhendislik Dergisi in Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi
- Vol. 17 (3)
A media-type-aware asynchronous framework for restoring heterogeneous stream endpoint collections
User-maintained Internet radio and IPTV-like playlists often contain unavailable, duplicate, redirected, malformed, or non-media endpoints. This article evaluates a reproducible, provenance-preserving Python framework for restoring heterogeneous stream endpoint collections through playlist-level media-type-aware validation, URL transformation, invalid-entry removal, and verified alternative selection. The framework parses TXT, M3U/M3U8, pipe-delimited, and AIMPPL4 inputs and normalizes station–URL pairs before applying asynchronous HTTP HEAD probing with a GET fallback whose response-body read is capped at one byte after HEAD 405/501. It then revalidates transformed URLs, selects one validated alternative per station, and exports machine-readable evidence. Across three repetitions on a fixed 400-endpoint sample in the primary cloud environment, a concurrency limit of 50 reduced mean validation time from 211.6 to 6.3 seconds, a 33.6-fold speedup. An independent repetition in a second environment showed the same timing direction. A separate full-corpus evaluation processed all 1,858 unique pairs in 31.14 seconds, validated 485 endpoints, and selected endpoints for 372 stations. In a four-configuration descriptive acceptance/repair analysis, the status-only/no-repair setting accepted 25.6 non-media 2xx responses on average, whereas the media-type/repair setting accepted 115.8 media-compliant endpoints and rescued 29.3. Within the sampled decoder strata, all 19 unique rescued URLs passed a short independent media probe and none of the 18 unique status-only non-media URLs did, although three of 20 validator-rejected URLs also passed. These results characterize time-bounded restoration of endpoint collections under the specified experimental conditions and do not establish recovery of remote services, sustained playback, or quality of experience.
- 6 July 2026
- Published by Turkish Journal of Engineering in Turkish Journal of Engineering
- Vol. 10 (3) , 919-935
Real-Time Bolt Defect Detection with Artificial Intelligence in Industrial Quality Control Applications: Comparative Evaluation of YOLO Architectures
This study presents an artificial intelligence-based system for the real-time and high-accuracy detection of defects in bolts moving on a production line conveyor belt. Bolt defects are critical for product safety and production quality, while manual inspection methods remain insufficient due to their time-consuming nature, susceptibility to human error, and limited reliability. Although various approaches have been proposed in the literature, existing methods are limited in terms of small-object detection, real-time performance, and industrial integration. To address these limitations, a deep learning-based defect detection mechanism was developed and comparatively evaluated using the YOLOv9, YOLOv10, and YOLO11 architectures, with emphasis on low hardware requirements, ease of integration, and real-time capability. A custom dataset of 12,075 high-resolution conveyor belt bolt images was constructed and used to train and validate all model variants. Experimental results demonstrate that all three architectures achieved over 99% mAP50, with the YOLO11x model reaching 99.39% mAP50 and 90.90% mAP50-95, closely matched by YOLOv10x at 99.39% mAP50 and 90.91% mAP50-95. Considering both detection accuracy and inference speed, YOLO11x provides the most favorable balance for production line integration, while YOLOv10x stands out as a competitive alternative with comparable accuracy and lower computational cost. The proposed system offers a faster, more reliable, and easily integrable solution than manual inspection, with strong potential for adoption in high-precision industrial sectors such as automotive, aerospace, and defense.
- 6 July 2026
- Published by Turkish Journal of Engineering in Turkish Journal of Engineering
- Vol. 10 (3) , 919-935
Real-Time Bolt Defect Detection with Artificial Intelligence in Industrial Quality Control Applications: Comparative Evaluation of YOLO Architectures
This study presents an artificial intelligence-based system for the real-time and high-accuracy detection of defects in bolts moving on a production line conveyor belt. Bolt defects are critical for product safety and production quality, while manual inspection methods remain insufficient due to their time-consuming nature, susceptibility to human error, and limited reliability. Although various approaches have been proposed in the literature, existing methods are limited in terms of small-object detection, real-time performance, and industrial integration. To address these limitations, a deep learning-based defect detection mechanism was developed and comparatively evaluated using the YOLOv9, YOLOv10, and YOLO11 architectures, with emphasis on low hardware requirements, ease of integration, and real-time capability. A custom dataset of 12,075 high-resolution conveyor belt bolt images was constructed and used to train and validate all model variants. Experimental results demonstrate that all three architectures achieved over 99% mAP50, with the YOLO11x model reaching 99.39% mAP50 and 90.90% mAP50-95, closely matched by YOLOv10x at 99.39% mAP50 and 90.91% mAP50-95. Considering both detection accuracy and inference speed, YOLO11x provides the most favorable balance for production line integration, while YOLOv10x stands out as a competitive alternative with comparable accuracy and lower computational cost. The proposed system offers a faster, more reliable, and easily integrable solution than manual inspection, with strong potential for adoption in high-precision industrial sectors such as automotive, aerospace, and defense.
- Open Access
- 30 June 2026
- Published by Sakarya University Journal of Computer and Information Sciences in Sakarya University Journal of Computer and Information Sciences
- Vol. 9 (3) , 868-882
Software Workforce Dynamics and Organizational Restructuring in Higher Education: A Retrospective Scenario Analysis
Large-scale software projects in universities often fall short due to excessive workload, insufficient staffing, and low motivation. Locating project-based software teams within the same organizational structure as administrative units leads to structural misalignments and intensifies these problems. This study examines workforce- and organization-related challenges through a comparative scenario analysis of a restructuring model proposed in 2006 at the IT Department of Sakarya University. The retrospective dataset includes evaluations of eight workforce factors provided by 27 staff members from two software development units and three administrative service units, reflecting participants’ perceptions of the scenario at the time and its anticipated impacts. Quantitative data were analyzed using statistical methods, while qualitative data were examined through thematic analysis. The findings indicate that software units experienced severe bottlenecks due to workload imbalance and staff shortages, whereas administrative units exhibited relatively balanced task distribution. According to the proposed scenario, relocating software teams to a project-oriented research center would increase workload acceptability by 43.1\% and staff adequacy by 44.6\% within the remaining administrative structure. The results suggest that this restructuring would substantially alleviate chronic workforce-related problems observed in software units. By interpreting historical data within its original context, this study demonstrates that past organizational projections can inform current debates on organizational transformation in higher education.
- 30 June 2026
- Published by Sakarya University Journal of Computer and Information Sciences in Sakarya University Journal of Computer and Information Sciences
- Vol. 9 (3) , 868-882
Open AccessSoftware Workforce Dynamics and Organizational Restructuring in Higher Education: A Retrospective Scenario Analysis
Large-scale software projects in universities often fall short due to excessive workload, insufficient staffing, and low motivation. Locating project-based software teams within the same organizational structure as administrative units leads to structural misalignments and intensifies these problems. This study examines workforce- and organization-related challenges through a comparative scenario analysis of a restructuring model proposed in 2006 at the IT Department of Sakarya University. The retrospective dataset includes evaluations of eight workforce factors provided by 27 staff members from two software development units and three administrative service units, reflecting participants’ perceptions of the scenario at the time and its anticipated impacts. Quantitative data were analyzed using statistical methods, while qualitative data were examined through thematic analysis. The findings indicate that software units experienced severe bottlenecks due to workload imbalance and staff shortages, whereas administrative units exhibited relatively balanced task distribution. According to the proposed scenario, relocating software teams to a project-oriented research center would increase workload acceptability by 43.1\% and staff adequacy by 44.6\% within the remaining administrative structure. The results suggest that this restructuring would substantially alleviate chronic workforce-related problems observed in software units. By interpreting historical data within its original context, this study demonstrates that past organizational projections can inform current debates on organizational transformation in higher education.
- Open Access
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
When Natural Bag Context Helps Wearable IoT Activity Recognition: Set-Conditioned Instance Prediction under Subject-Local Semantics
Wearable Internet of Things (IoT) systems rely on window-level activity recognition for cyber-physical-human sensing. Most pipelines classify each sensing window independently, even though windows are naturally organized by subject, device, session, or short-term behavioral regime. This paper studies whether instance-level activity prediction can be improved by conditioning each window on its natural bag context while preserving one prediction per window. The evaluation spans five public sensing benchmarks (HAR, HAPT, MHEALTH, PAMAP2, WISDM) and compares independent neural and XGBoost baselines, grouped-feature baselines, and a permutation-invariant DeepSet Context model under natural bags, within-subject local bags, and size-matched random-bag controls, across 10 seeds and bag sizes {4,8,16,32}, with accuracy as the primary metric. Natural context is useful but regime-dependent. On HAPT under subject-full grouping, the DeepSet Context model improves accuracy from 0.9321 to 0.9450 (a +1.29 percentage-point gain over the strongest independent baseline; paired t-test p<0.001), with a natural-specific gain of +1.15 percentage points after random-bag adjustment; HAR is supportive but mixed, MHEALTH and PAMAP2 are dominated by tabular or grouped-feature baselines, and WISDM remains marginal and unsupported by the timing benchmark. A representative timing benchmark also shows a non-uniform IoT cost-benefit profile: context is both more accurate and faster on HAR, more accurate but slower on HAPT, and not justified on MHEALTH, PAMAP2, or WISDM. Learned natural context should therefore be used when bags encode coherent subject-local structure and peer windows add information beyond the target window, while random-bag controls and grouped-feature baselines remain essential before attributing gains to meaningful context.
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWhen Natural Bag Context Helps Wearable IoT Activity Recognition: Set-Conditioned Instance Prediction under Subject-Local Semantics
Wearable Internet of Things (IoT) systems rely on window-level activity recognition for cyber-physical-human sensing. Most pipelines classify each sensing window independently, even though windows are naturally organized by subject, device, session, or short-term behavioral regime. This paper studies whether instance-level activity prediction can be improved by conditioning each window on its natural bag context while preserving one prediction per window. The evaluation spans five public sensing benchmarks (HAR, HAPT, MHEALTH, PAMAP2, WISDM) and compares independent neural and XGBoost baselines, grouped-feature baselines, and a permutation-invariant DeepSet Context model under natural bags, within-subject local bags, and size-matched random-bag controls, across 10 seeds and bag sizes {4,8,16,32}, with accuracy as the primary metric. Natural context is useful but regime-dependent. On HAPT under subject-full grouping, the DeepSet Context model improves accuracy from 0.9321 to 0.9450 (a +1.29 percentage-point gain over the strongest independent baseline; paired t-test p<0.001), with a natural-specific gain of +1.15 percentage points after random-bag adjustment; HAR is supportive but mixed, MHEALTH and PAMAP2 are dominated by tabular or grouped-feature baselines, and WISDM remains marginal and unsupported by the timing benchmark. A representative timing benchmark also shows a non-uniform IoT cost-benefit profile: context is both more accurate and faster on HAR, more accurate but slower on HAPT, and not justified on MHEALTH, PAMAP2, or WISDM. Learned natural context should therefore be used when bags encode coherent subject-local structure and peer windows add information beyond the target window, while random-bag controls and grouped-feature baselines remain essential before attributing gains to meaningful context.
- Open Access
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
An Interpretable CF-RL-TOPSIS Fusion Model for Skills-Aware Talent Recommendation
Skills-aware talent recommendation draws on three families of signals: collaborative behavioral regularities, adaptive trajectory dynamics, and interpretable occupation-side criteria; yet public benchmark evidence on their complementarity is scarce. This study introduces CF-RL-TOPSIS, an interpretable late-fusion architecture that combines a transition-aware collaborative branch, a lightweight reinforcement-style bandit over occupation families, and an entropy-weighted Technique for Order Preference by Similarity to Ideal Solution branch built from six semantic proxies, with validation-selected fusion weights that remain directly inspectable. Evaluation proceeds on two frozen public information and communication technology talent-history benchmarks, JobHop and Karrierewege, under repeated chronological top-5 ranking with paired Wilcoxon testing. On JobHop the full hybrid attains NDCG@5 = 0.3040 ± 0.0073 and significantly surpasses repeat-last, item Markov, transition-aware collaborative filtering, the CF+TOPSIS hybrid, GRU4Rec, and SASRec (p ? 0.0039 across planned comparisons). On Karrierewege the hybrid remains competitive but does not significantly exceed the strongest Markov baseline, exposing a persistence-dominated regime in which the bandit branch correctly collapses to near-zero weight. Proxy-sensitivity, family-level deep Q-network, and runtime checks corroborate this reading, and a worked user-level case illustrates branch-score, criterion-weight, and rank-shift inspection for an individual recommendation. The contribution is therefore not a benchmark-agnostic superiority claim but a reproducible account of when an interpretable late-fusion architecture meaningfully complements simple continuation heuristics. In semantically rich, non-saturating talent-history regimes the three branches reinforce one another; in persistence-dominated regimes the same architecture remains competitive through its collaborative backbone, with the adaptive branch correctly inactive.
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAn Interpretable CF-RL-TOPSIS Fusion Model for Skills-Aware Talent Recommendation
Skills-aware talent recommendation draws on three families of signals: collaborative behavioral regularities, adaptive trajectory dynamics, and interpretable occupation-side criteria; yet public benchmark evidence on their complementarity is scarce. This study introduces CF-RL-TOPSIS, an interpretable late-fusion architecture that combines a transition-aware collaborative branch, a lightweight reinforcement-style bandit over occupation families, and an entropy-weighted Technique for Order Preference by Similarity to Ideal Solution branch built from six semantic proxies, with validation-selected fusion weights that remain directly inspectable. Evaluation proceeds on two frozen public information and communication technology talent-history benchmarks, JobHop and Karrierewege, under repeated chronological top-5 ranking with paired Wilcoxon testing. On JobHop the full hybrid attains NDCG@5 = 0.3040 ± 0.0073 and significantly surpasses repeat-last, item Markov, transition-aware collaborative filtering, the CF+TOPSIS hybrid, GRU4Rec, and SASRec (p ? 0.0039 across planned comparisons). On Karrierewege the hybrid remains competitive but does not significantly exceed the strongest Markov baseline, exposing a persistence-dominated regime in which the bandit branch correctly collapses to near-zero weight. Proxy-sensitivity, family-level deep Q-network, and runtime checks corroborate this reading, and a worked user-level case illustrates branch-score, criterion-weight, and rank-shift inspection for an individual recommendation. The contribution is therefore not a benchmark-agnostic superiority claim but a reproducible account of when an interpretable late-fusion architecture meaningfully complements simple continuation heuristics. In semantically rich, non-saturating talent-history regimes the three branches reinforce one another; in persistence-dominated regimes the same architecture remains competitive through its collaborative backbone, with the adaptive branch correctly inactive.
- Open Access
- 10 June 2026
- Published by Elsevier in SSRN Electronic Journal
Branching-Controlled Disentanglement of Dendritic Nonlinearity from Routing and Capacity in Parameter-Matched Image Classification
Dendritic artificial neural networks (DANNs) add branch-level local computation before somatic integration, yet reported gains often conflate three sources of advantage: dendritic nonlinearity, routing structure, and parameter capacity. This work introduces a branching-controlled isolation framework whose methodological core is the Naive-Branch control, a baseline that inherits the routing tensor and parameter shapes of DANN-LRF but omits the dendrite-level nonlinearity, so any gap is attributable to the per-branch nonlinearity rather than to routing or capacity. Together with a parameter-matched multilayer perceptron (MLP-Param) and a higher-capacity dense reference (VANN-Same), this triad disentangles all three sources, and the primary metric is reconstructed at the validation-selected epoch from per-epoch histories to remove test-selection bias, with a sensitivity audit confirming that condition leaders are preserved under both metric definitions. Evaluations span flattened FashionMNIST, KMNIST, and CIFAR-10, with fixed-test reduced training-data conditions and a CPU-only timing protocol. On full data, DANN-LRF outperforms MLP-Param by 1.67, 4.06, and 17.04 percentage points under paired Wilcoxon tests with Holm correction, and exceeds Naive-Branch on FashionMNIST and CIFAR-10 but not on KMNIST, where DANN-RANDOM is numerically best; under reduced training data, the DANN-LRF - MLP-Param advantage is preserved and DANN-LRF adds only 1% to 4% runtime overhead beyond Naive-Branch. The disentanglement yields three findings: dendritic nonlinearity contributes a small but measurable increment beyond routing, the preferred dendritic sampling strategy is dataset-dependent rather than universal, and DANNs are most defensible for compact classification under matched budgets when structured feature interactions are plausible.
- 10 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBranching-Controlled Disentanglement of Dendritic Nonlinearity from Routing and Capacity in Parameter-Matched Image Classification
Dendritic artificial neural networks (DANNs) add branch-level local computation before somatic integration, yet reported gains often conflate three sources of advantage: dendritic nonlinearity, routing structure, and parameter capacity. This work introduces a branching-controlled isolation framework whose methodological core is the Naive-Branch control, a baseline that inherits the routing tensor and parameter shapes of DANN-LRF but omits the dendrite-level nonlinearity, so any gap is attributable to the per-branch nonlinearity rather than to routing or capacity. Together with a parameter-matched multilayer perceptron (MLP-Param) and a higher-capacity dense reference (VANN-Same), this triad disentangles all three sources, and the primary metric is reconstructed at the validation-selected epoch from per-epoch histories to remove test-selection bias, with a sensitivity audit confirming that condition leaders are preserved under both metric definitions. Evaluations span flattened FashionMNIST, KMNIST, and CIFAR-10, with fixed-test reduced training-data conditions and a CPU-only timing protocol. On full data, DANN-LRF outperforms MLP-Param by 1.67, 4.06, and 17.04 percentage points under paired Wilcoxon tests with Holm correction, and exceeds Naive-Branch on FashionMNIST and CIFAR-10 but not on KMNIST, where DANN-RANDOM is numerically best; under reduced training data, the DANN-LRF - MLP-Param advantage is preserved and DANN-LRF adds only 1% to 4% runtime overhead beyond Naive-Branch. The disentanglement yields three findings: dendritic nonlinearity contributes a small but measurable increment beyond routing, the preferred dendritic sampling strategy is dataset-dependent rather than universal, and DANNs are most defensible for compact classification under matched budgets when structured feature interactions are plausible.
- Open Access
- 19 May 2026
- Published by Elsevier in SSRN Electronic Journal
How can criterion-rich e-marketplaces recommend without CF? An explainable hybrid RL and entropy-weighted TOPSIS framework
E-marketplaces increasingly operate with rich product-level commercial criteria, such as price, rating, review density, and category structure, while their collaborative interaction signals remain sparse, restricted by privacy regulation, or intentionally excluded from the deployable architecture. Regulatory pressure under the European Union Digital Services Act and the forthcoming AI Act, together with platform-governance limits, is turning this collaborative-filtering-free (CF-free) regime into a deployment constraint that operators inherit rather than choose, while requiring that ranking decisions remain auditable. This paper proposes Hybrid RL-TOPSIS, a CF-free decision-support framework coupling an entropy-weighted Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) ranker with a tabular Q-learning behavioral module through a transparent linear fusion rule, so that every recommendation decomposes into a structural and a behavioral contribution. Across 50 deterministic 400-item Amazon India bootstrap catalogs, Hybrid RL-TOPSIS reaches F1@7 of 0.901 versus 0.574 for RL-only and 0.259 for TOPSIS-only, with paired significance tests confirming both gaps (Delta F1 = +0.326, p = 5.24 x 10^-37; Delta F1 = +0.641, p = 6.38 x 10^-51). The advantage persists under reduced reward shaping and gradual concept drift. Boundary checks on the McAuley Home and Kitchen 5-core dataset, sparse collaborative benchmarks, CUDA-based deep recommenders, LinUCB, and varying catalog sizes identify when interaction-rich architectures remain preferable, defining the operational niche rather than claiming universal superiority. A SHAP audit of hybrid scores and intrinsic RL-versus-TOPSIS contribution tables make the framework auditable at criterion level for managerial review and regulatory disclosure under tight compute budgets (16 KB Q-table, no GPU, sub-millisecond inference).
- 19 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHow can criterion-rich e-marketplaces recommend without CF? An explainable hybrid RL and entropy-weighted TOPSIS framework
E-marketplaces increasingly operate with rich product-level commercial criteria, such as price, rating, review density, and category structure, while their collaborative interaction signals remain sparse, restricted by privacy regulation, or intentionally excluded from the deployable architecture. Regulatory pressure under the European Union Digital Services Act and the forthcoming AI Act, together with platform-governance limits, is turning this collaborative-filtering-free (CF-free) regime into a deployment constraint that operators inherit rather than choose, while requiring that ranking decisions remain auditable. This paper proposes Hybrid RL-TOPSIS, a CF-free decision-support framework coupling an entropy-weighted Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) ranker with a tabular Q-learning behavioral module through a transparent linear fusion rule, so that every recommendation decomposes into a structural and a behavioral contribution. Across 50 deterministic 400-item Amazon India bootstrap catalogs, Hybrid RL-TOPSIS reaches F1@7 of 0.901 versus 0.574 for RL-only and 0.259 for TOPSIS-only, with paired significance tests confirming both gaps (Delta F1 = +0.326, p = 5.24 x 10^-37; Delta F1 = +0.641, p = 6.38 x 10^-51). The advantage persists under reduced reward shaping and gradual concept drift. Boundary checks on the McAuley Home and Kitchen 5-core dataset, sparse collaborative benchmarks, CUDA-based deep recommenders, LinUCB, and varying catalog sizes identify when interaction-rich architectures remain preferable, defining the operational niche rather than claiming universal superiority. A SHAP audit of hybrid scores and intrinsic RL-versus-TOPSIS contribution tables make the framework auditable at criterion level for managerial review and regulatory disclosure under tight compute budgets (16 KB Q-table, no GPU, sub-millisecond inference).
- 2 May 2026
- Published by IGI Global Scientific Publishing
- p. 203-230
Zero Trust Architecture and Enterprise Applications
As industry grows more digital, securing IoT systems in hybrid cloud environments is vital. This chapter presents Zero Trust Architecture (ZTA), which treats every user, device and network element as untrusted by default. Access requests are evaluated continuously based on identity, device health and context. Core principles include multi-factor authentication, least-privilege access, continuous monitoring, micro-segmentation and risk-based decisions. These help protect cloud services, IoT devices, remote workers and microservices. NIST Special Publication 800-207 defines policy enforcement point, policy administrator and policy engine. The chapter reviews implementations such as Google BeyondCorp, Cisco ISE and AWS GuardDuty, and discusses challenges like resistance to change, technical gaps, user experience and cost. Despite these hurdles, ZTA can reduce breaches, improve visibility and ease compliance. Supported by maturity models and case studies, this chapter offers a concise review of ZTA's foundations and benefits.
- 2 May 2026
- Published by IGI Global Scientific Publishing
- p. 203-230
Zero Trust Architecture and Enterprise Applications
As industry grows more digital, securing IoT systems in hybrid cloud environments is vital. This chapter presents Zero Trust Architecture (ZTA), which treats every user, device and network element as untrusted by default. Access requests are evaluated continuously based on identity, device health and context. Core principles include multi-factor authentication, least-privilege access, continuous monitoring, micro-segmentation and risk-based decisions. These help protect cloud services, IoT devices, remote workers and microservices. NIST Special Publication 800-207 defines policy enforcement point, policy administrator and policy engine. The chapter reviews implementations such as Google BeyondCorp, Cisco ISE and AWS GuardDuty, and discusses challenges like resistance to change, technical gaps, user experience and cost. Despite these hurdles, ZTA can reduce breaches, improve visibility and ease compliance. Supported by maturity models and case studies, this chapter offers a concise review of ZTA's foundations and benefits.
- Open Access
- 6 March 2026
- Published by Elsevier in SSRN Electronic Journal
Stabilizing Policy Convergence in 3D Container Loading: A Comparative Study of Classical and Quantum-Enhanced Reinforcement Learning
The three-dimensional container loading problem in maritime logistics constitutes a high-dimensional combinatorial optimization challenge where classical reinforcement learning (RL) algorithms often exhibit learning instability as problem complexity scales. This paper investigates whether embedding parameterized quantum circuits (PQCs) into the policy networks of Proximal Policy Optimization (PPO) and Asynchronous Advantage Actor-Critic (A3C) can stabilize policy convergence under increasing spatial complexity. A systematic scale-based experimental framework is designed with 400, 600, 800, and 1000 box scenarios, each evaluated across 5, 10, and 15 qubit configurations using five independent random seeds. Results show that all models converge deterministically at the first episode for scenarios up to 800 boxes, with zero cross-seed variance. At the 1000-box threshold, classical models begin to exhibit stochastic convergence and seed sensitivity; the classical A3C, for instance, produces a cross-seed AUC standard deviation of 7.46. In contrast, the 15-qubit QPPO configuration reduces this deviation to 1.80, representing a 71.3% improvement in learning consistency without sacrificing fill rate. Visual placement analysis further confirms that higher qubit configurations generate more compact and structurally organized spatial layouts with reduced micro-gap density. These findings indicate that the primary benefit of quantum-enhanced RL in dense packing optimization is not raw performance gain but improved policy convergence stability under severe combinatorial complexity.
- 6 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessStabilizing Policy Convergence in 3D Container Loading: A Comparative Study of Classical and Quantum-Enhanced Reinforcement Learning
The three-dimensional container loading problem in maritime logistics constitutes a high-dimensional combinatorial optimization challenge where classical reinforcement learning (RL) algorithms often exhibit learning instability as problem complexity scales. This paper investigates whether embedding parameterized quantum circuits (PQCs) into the policy networks of Proximal Policy Optimization (PPO) and Asynchronous Advantage Actor-Critic (A3C) can stabilize policy convergence under increasing spatial complexity. A systematic scale-based experimental framework is designed with 400, 600, 800, and 1000 box scenarios, each evaluated across 5, 10, and 15 qubit configurations using five independent random seeds. Results show that all models converge deterministically at the first episode for scenarios up to 800 boxes, with zero cross-seed variance. At the 1000-box threshold, classical models begin to exhibit stochastic convergence and seed sensitivity; the classical A3C, for instance, produces a cross-seed AUC standard deviation of 7.46. In contrast, the 15-qubit QPPO configuration reduces this deviation to 1.80, representing a 71.3% improvement in learning consistency without sacrificing fill rate. Visual placement analysis further confirms that higher qubit configurations generate more compact and structurally organized spatial layouts with reduced micro-gap density. These findings indicate that the primary benefit of quantum-enhanced RL in dense packing optimization is not raw performance gain but improved policy convergence stability under severe combinatorial complexity.