A Puma Optimized Reinforcement Learning Model for Detection of Results Anomalies in Higher Education

Examination results are crucial educational assets that must be carefully processed, stored and protected, because the integrity of the eventual certificate depends largely on the accuracy and validity of the results. Students’ admission rate into higher institution in Nigeria increases yearly, making it challenging for the available limited manpower and existing legacy infrastructure to contain the magnitude of irregularities in results processing. This usually cause delay in approving results for further decision making. Also, educational institutions in Nigeria frequently encounters unwholesome results manipulations, leading to certificate forgery. Considering the complexity and crucial nature of the activities involved in conducting end of semester examinations, appropriate technology must be deployed to ensure the accuracy and integrity of the process. Building anomaly detection into result computation process becomes necessary, if the problem of certificate forgery is to be solved. This study investigated the combination of machine learning and blockchain technologies in developing a secure, intelligent, and integrated system for results filtration and storage respectively. With design, training, and testing simulated in MATLAB, traditional security features of the blockchain technology, particularly its immutable and distributed ledger, were combined with the intelligence and predictive capabilities of a Puma optimized reinforcement learning agent in a Q-Learning architecture, to build a real-time results filtering and storage platform. Its convergence speed and global search ability was significantly enhanced, leading to more optimal selection of key reinforcement learning hyperparameters that are critical to the effectiveness of RL-based detection models. Demonstrating the advantage of integrating Puma Optimizer into the conventional RL, the Puma Optimized Reinforcement Learning model returned superior performance across all evaluation metrics. The effectiveness of the model was measured by comparing its performance with those of traditional reinforcement learning model, using suitable metrics. With a specificity of 99.53%, Precision of 98.11%, Accuracy of 99.01%, and Computation Time of 42.38s over 800 epochs; the model was considered highly appropriate for real-time detection of results anomalies.