The scheduling of unrelated parallel machines (UPM) is a critical aspect of production planning and a major challenge in the manufacturing industry. Scheduling outcomes directly impact a business's profitability and sustainability. In the UPM problem, the processing times for each work order vary across machines and are not proportionally consistent. This becomes a common scheduling challenge for medium- to large-sized enterprises due to its complexity. To address the UPM scheduling problem, this study aims to minimize the makespan while considering sequence-dependent setup times for both machines and work orders. It then proposes a genetic algorithm integrating reinforcement learning and variable neighborhood search. To assess the feasibility and effectiveness of the proposed algorithm, real work orders from a renowned international semiconductor copper foil manufacturer (Company D) are used as test data to assess the proposed method. The results indicate that the proposed method can reduce the makespan by 22.45% compared to other existing methods used by Company D. Additionally, randomly generated test data were used to validate the effectiveness and feasibility of the method.

This study presents a Reinforcement Learning (RL)-assisted Genetic Algorithm (GA) framework for the Multi-Objective Capacitated Vehicle Routing Problem with Time Windows (MOCVRPTW). In this problem, a set of homogeneous vehicles depart from a depot, visit all customers exactly once, and return back to the depot. The routes of the vehicles are constructed by considering three objectives: minimizing the total travel time, minimizing the number of vehicles, and maximizing the satisfaction obtained from the customers who are visited within their time windows. We propose using an NSGA-II-based approach that is assisted by Q-learning-based operator selection methods for this problem. Unlike traditional GAs that use fixed operators, the proposed approach enables learning-based selection of each operator (crossover and mutation) considering the current performance of solutions. We make tests with five different Q-learning-based operator selection strategies and compare their results to using fixed or randomly selected operators by nonparametric statistical methods. The results show that all Q-learning-based operator selection strategies outperform the fixed-operator approach, whereas the random selection strategy is outperformed by four. In addition, when the best operator for each state of solutions is found considering all solution approaches and used in NSGA-II throughout the algorithm, it consistently results in the best performance among all. Overall, the results demonstrate that the proposed RL-supported GA framework provides a competitive alternative in terms of Pareto-front solution quality for MOCVRPTW, and learning-based operator selection can be an effective mechanism for adaptively controlling the evolutionary search process.

The maximum tardiness reflects the worst level of service associated with customer needs; thus, the principle that seru production reduces the maximum tardiness is investigated, and a model to minimize the maximum tardiness of the seru production system is established. In order to obtain the exact solution, the non-linear seru production model with minimizing the maximum tardiness is split into a seru formation model and a linear seru scheduling model. We propose an efficient cooperative algorithm using a genetic algorithm and an innovative reinforcement learning algorithm (CAGARL) for large-scale problems. Specifically, the GA is designed for the seru formation problem. Moreover, the QL-seru algorithm (QLSA) is designed for the seru scheduling problem by combining the features of meta-heuristics and reinforcement learning. In the QLSA, we design an innovative QL-seru table and two state trimming rules to save computational time. After extensive experiments, compared with the previous algorithm, CAGARL improved by an average of 56.6%. Finally, several managerial insights on reducing maximum tardiness are proposed.

Experience plays a vital role in the decision-making (DM) process. In this paper simulation, modeling, and analysis of past experience over DM has been done using the Iowa gambling task (IGT). The Human DM process is very complex and difficult to model through computational methods because it is a subjective type of process and varies person-to-person. Therefore, this study is an attempt to simulate a DM model similar to the human DM process. For this collection of real data was done and was provided as input to the developed eight Reinforcement Learning (RL) models. The result shows that the performance of the model based on Prospect Utility (PU) learned with Decay Reinforcement Rule (DRI) and Trial Dependency Choice (TDC) is better as compared to other models. It is observed from the analysis of data and also validated that simulation and models output that the experienced group performs better than inexperienced.

Dynamic Flexible Job Shop Scheduling (DFJSS) is a critical challenge in smart manufacturing due to dynamic job arrivals, machine breakdowns, routing flexibility, and multi-objective performance requirements. Conventional dispatching rules and metaheuristic algorithms often fail to provide adaptive real-time decisions. This paper proposes a Greedy Actor Neural Critic PPO based Reinforcement Learning (GANC-PPO-RL) framework to minimize makespan and total energy consumption. The problem is formulated as a Markov Decision Process, where machine availability, job progress, queue status, processing times, and energy levels define states, and job–machine assignments define actions. A multi-objective reward penalizes higher makespan and energy usage. The PPO algorithm ensures stable policy updates using a clipped objective. Results show makespan within ±2% of benchmarks for small instances and about 23% reduction for larger cases, with 16–20% lower makespan and 15–25% lower energy consumption compared to RL-QL.

The rapid advancement of artificial intelligence (AI) has revolutionized project portfolio management (PPM), as it has in many other areas, by introducing data-driven methods that improve decision-making, risk assessment, and strategic alignment. Unlike traditional project management, which emphasizes individual project execution, PPM requires balancing multiple initiatives to optimize value creation and resource allocation. This paper presents a systematic review of scientific research on the integration of AI techniques into PPM, focusing on their applications, benefits, and challenges. The review synthesizes findings from 73 peer-reviewed studies covering a wide range of AI methodologies, such as machine learning, deep learning, neural networks, reinforcement learning, natural language processing, and hybrid optimization models. These approaches have been applied in diverse fields, including information technology, construction, healthcare, defense, energy, and telecommunications. Analysis shows that AI significantly improves project portfolio performance by predicting project outcomes, identifying interdependencies, optimizing resource allocation, and supporting adaptive strategies in dynamic environments. In addition, advanced AI tools provide project portfolio managers with predictive and prescriptive analytics, transforming PPM from reactive monitoring to proactive governance. Despite these advances, challenges remain regarding data quality, organizational readiness, and interpretability of AI-based models. Concerns about transparency, ethical implications, and integration with existing management frameworks also hinder wider adoption. However, recent developments indicate a growing trend toward hybrid systems that combine AI with traditional decision-making models, increasing both accuracy and practical applicability. This review contributes to theory and practice by synthesizing current knowledge, highlighting research gaps, and identifying emerging directions such as the use of large language models, ensemble methods, and sustainability-focused project portfolio optimization. The findings highlight the transformative potential of AI in advancing PPM and provide valuable insights for researchers and practitioners seeking to design smarter, more adaptive, and more sustainable project portfolio management strategies.

The quick transfer of academic processes to cloud-based systems has increased the cyber-attack area of academic institutions significantly by introducing systemic weaknesses into the learning management systems (LMS), enterprise resource planning (ERP) tools, as well as databases housing student data. Traditional security controls, relying mostly on fixed signature databases and rule-based intrusion detection, are no longer effective against polymorphic malware, zero-day attacks, and advanced insider threats that define contemporary adversarial environments. Nevertheless, even with the emergence of cloud-native security tools, educational organisations are experiencing a high level of false positives, slower response to threats, and an almost complete inability to predict new vectors of attacks. Lacking contextually sensitive self-updating intelligence pipes exposes campus networks to cascading failures that threaten the privacy of students, integrity of research data, and compliance with regulations. This paper introduces Adaptive Threat Intelligence Learning Algorithm (ATILA), a five-layer security architecture, which combines transformer-based anomaly detection, reinforcement-learning-based policy coordination and federated collaborative defence in geographically distributed institutional nodes. Three benchmark datasets were used to evaluate ATILA: UNSW-NB15, CICIDS-2018 and a simulated corpus of campus cloud logs simulating a federated multi-campus environment. They were compared to rule-based, intrusion detection systems, and non-adaptive deep learning models and rule-based classifiers based on the static random forest. ATILA had 98.3% and 1.4% detection and false positive rates, respectively, which is 4.6 and 8.3 % points better than the most successful non-adaptive baseline. The recall of the zero-day detection went to 91.7 and the mean time of the threat response came down to 6.1 ms. University LMS environments, multi-campus cloud federations, and online examination systems all receive deployment ready design specifications that provide a repeatable pathway towards institutions wishing to operationalise adaptive AI-driven security at scale.

The rise of Internet of Things (IoT) technologies in smart solar energy systems has presented a serious cybersecurity issue, especially with the close interdependence between cyber infrastructure and physical functions. Traditional intrusion detection systems (IDS) are mainly based on the network level of abnormalities and do not consider the system level effects leading to poor mitigation. In this paper, we suggest an AI-based cyber-physical cybersecurity architecture, which combines an intrusion detection model in the form of CNNLSTM and a mitigation mechanism based on reinforcement learning. A decision fusion layer is added to integrate cyber threats scores with photovoltaic system conditions to respond in context. Experimental analysis on real-world data sets indicates that the framework proposed has a 98.3% detection rate, 0.97 F1-score, and a lower false positive of 0.02, which performs better than the current approaches by up to 2.1% accuracy. The framework enhances energy stability at the system level up to 96.8% and recovery time is shortened by about 58 percent when the system is under attack. Moreover, the model has a strong performance of more than 96% accuracy even when there is a high attack intensity situation. These results highlight the value of cyber intelligence and integration of physical system awareness. The proposed framework is holistic and realistic to provide a guarantee of the safety of smart solar energy systems to ensure not only high detection but also stability of the operations in real-life scenarios.

Autonomous ground robots are increasingly relied upon in dynamic environments where reliable navigation and context-aware interaction are essential. However, existing robotic control systems often rely on cloud-based reinforcement learning (RL) frameworks or static algorithms that fail to adapt in real-time to noisy, unpredictable scenarios. These models typically overlook the constraints of edge deployment and lack robust integration with human interaction modalities such as voice and semantic object awareness. To address these limitations, this work proposes a fully embedded, IoT-enabled ground robot powered by a reinforcement learning-based adaptive control framework. The system leverages Raspberry Pi 4B+ as its core computational unit, integrating MQTT-driven communication, multimodal interaction through speech and vision, and lightweight policy convergence for obstacle-aware navigation. A novel RL-based state-action pipeline is trained and deployed entirely on-device, ensuring real-time responsiveness without external computation. Experimental evaluations show that the proposed framework reduces navigation errors by 22% and improves interaction latency by 37% over traditional PID and A*-based systems. The RL model converges in under 2200 episodes, with stable reward curves and high reliability across variable acoustic and physical terrains. This study showcases how low-cost, edge-based robots can achieve high autonomy and situational awareness contributing to future advancements in resilient, self-adaptive robotic systems within smart and resource-constrained environments.

The effectiveness and efficiency of a machine learning model can be improved by feature selection, especially for high-dimensional datasets such as in cybersecurity. The proposed approach utilizes an enhanced version of the Rainbow agent with a memory storage structure. The suggested approach is assessed using two benchmark datasets namely RT-IoT2022 which is targeted towards IoT network security and the Android Malware Detection dataset which is meant for mobile security. The specification of the reinforcement learning model has been trained for 20 epochs and it is progressively enhanced through feature subsets to enhance classification accuracy. The results show that the AUC scores continuously increase were the one for RT-IoT2022 achieves 0.91 and Android at 0.93. Three well-known classifiers XGBoost, Random Forest and multi-layer perceptron (MLP) are used to test the power of the selected features. The outcome evaluation on RT-IoT2022 dataset shows that Random Forest achieved maximum accuracy (99.48%), followed by XGBoost (99.16%), while MLP secured 94.04% accuracy. In the Android malware dataset, XGBoost model gave the best accuracy of 89.50%, followed closely by Random Forest with 87.00% and MLP with 86.50%. This clearly shows that reinforcement learning based feature selection enhances accuracy and reduces computation. The research emphasizes utilizing dynamic feature selection in any cyber security application. The future will experiment with incorporating deep reinforcement learning as well as hybrid selection.