ENHANCED TASK PRIORITIZATION SYSTEM USING DEEP-Q-NETWORK MODEL

International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 113 13 ENHANCED TASK PRIORITIZATION SYSTEM USING DEEP-Q-NETWORK MODEL 1.Nnaemeka Kingsley Ugwumba Institution: University of Port Harcourt, Port Harcourt, Nigeria Email Address:[email protected] 2.Peter Sunday Jaja Institution: University of Port Harcourt, Port Harcourt, Nigeria Email Address:[email protected] Corresponding author: Nnaemeka Kingsley Ugwumba Abstract This study presents the development and evaluation of a Deep Q Network model for intelligent task prioritization, addressing the limitations of static, manual methods in traditional to do systems. The research utilizes a synthetically generated dataset of task attributes including deadlines, complexity, and priority scores created with Python's Pandas and NumPy libraries to simulate real world scenarios. This dataset enabled the training and validation of a reinforcement learning agent that autonomously learns optimal prioritization strategies based on user behavior and contextual factors. The proposed Deep Q Network model was evaluated against baseline methods including Earliest Deadline First and a Static Eisenhower Matrix, demonstrating superior performance with 92.3% prioritization accuracy and a 93.5% deadline adherence rate. The results highlight the significant potential of deep reinforcement learning for dynamic task management, providing a foundation for future integration into productivity tools through a proposed system architecture incorporating Django and React Native. Keywords: Artificial Intelligence, Task Prioritization, Deep Q-Network (DQN), Reinforcement Learning, Synthetic Data, Productivity Tools. Introduction Effective task management is critical for personal and professional productivity. Conventional to-do lists, however, are often static and rely on manual prioritization, a process that can be timeconsuming, inefficient, and poorly adapted to dynamic real-world contexts. The advancement of artificial intelligence (AI) presents an opportunity to overcome these limitations through the development of adaptive systems that can automatically prioritize tasks based on user behavior and contextual factors. International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 114 13 While previous approaches have incorporated rule-based systems or static priority matrices like the Eisenhower Box, they often lack the ability to learn and adapt to individual user patterns. Reinforcement Learning (RL), a subset of AI that enables agents to learn optimal behaviors through interaction with an environment, offers a promising framework for this challenge. This study, therefore, focuses on the development and validation of a reinforcement learningbased model for intelligent task prioritization. The primary contribution of this work is the design and experimental evaluation of a Deep Q-Network (DQN) agent that learns to prioritize tasks dynamically based on a multifaceted state representation, including deadlines, complexity, and user context. To validate the model, we generated a controlled simulated task dataset. Furthermore, we propose a full-stack system architecture to illustrate how this model can be integrated into a practical task management application. The key objectives of this research are: i. To design a DQN model with a state and reward structure tailored for the task prioritization problem. Ii .To generate and utilize a simulated task dataset for the training and initial validation of the proposed model. ii. To evaluate the model's performance against baseline prioritization methods in a simulated environment. iii. To propose a software system architecture for deploying the model in a real-world application Purpose and Objective of the Data The data presented in this study were collected to support the development of an AI-powered todo list application designed to address the limitations of traditional task management systems. Conventional tools provide a static framework for organizing tasks but lack the ability to dynamically adjust to new priorities, often leading to inefficiencies, missed deadlines, and increased stress. This project therefore sought to generate a dataset that would allow for the design and validation of a reinforcement learning model capable of adaptive prioritization. The dataset was created using Python libraries such as Pandas and NumPy to simulate realistic task scenarios. Variables included task identifiers, deadlines, urgency levels, task complexity, historical user behavior, and contextual information such as calendar events and user availability. These variables reflect the multifaceted nature of decision-making in task management and were chosen to enable the Deep Q-Network (DQN) to learn optimal prioritization strategies. Regression analysis was further applied to examine how these task attributes influenced productivity outcomes. The data have not been published as part of a previous research paper; they were generated exclusively for this project to test and validate the proposed model. Related studies in the International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 115 13 literature (Mnih et al., 2020; Sharma & Gupta, 2022; Zhang & Lee, 2023) informed the design of the dataset, but no prior dataset replicates this exact configuration. We believe that this dataset is valuable beyond the scope of this project. It provides a foundation for further research into intelligent task management, reinforcement learning applications in productivity tools, and adaptive scheduling systems. Future researchers could extend this dataset to collaborative settings or integrate it with real-world data streams such as emails and project management platforms. Data description The dataset acquired for this research consists of simulated task management data generated in Python using Pandas and NumPy libraries. The dataset was specifically designed to reflect realistic task management scenarios in order to train and evaluate a Deep Q-Network (DQN) model for intelligent task prioritization. The data include multiple task attributes: a. Task identifiers and descriptions (unique IDs and text-based task labels) b. Deadlines (expressed in date and time format) c. Urgency levels (categorized on a numerical scale) d. Task complexity ratings (quantitative values representing difficulty) e. Historical task behavior (completion times, delays, frequency of task type) f. Contextual variables (calendar events, availability windows, and workload constraints) These variables were designed to reflect the real-world diversity of user tasks and priorities. Together, they provide the foundation for reinforcement learning to optimize prioritization decisions. Data Processing The data were processed in several steps to ensure usability for machine learning. First, raw simulated task attributes were structured into tabular datasets. Missing values and inconsistencies were resolved programmatically. Categorical features, such as urgency levels, were encoded numerically to be compatible with the learning algorithms. Task completion times were normalized, while contextual variables (e.g., deadlines and events) were standardized for consistency across simulations. This preprocessing ensured that the DQN model could efficiently learn from the data without bias introduced by scale or representation differences. Methodological Notes International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 116 13 The data were generated through simulation rather than field collection, using Python scripts to create task scenarios that mimic real-world environments. The simulation incorporated randomized parameters to replicate variation in deadlines, urgency, and task dependencies. The data were subsequently stored in structured formats (CSV files) for model training and evaluation. No external hardware or experimental equipment was required, as the data were computationally produced. The dataset was then used to train and validate the AI model under controlled conditions, with separate data files created for training, testing, and evaluation. Each file was labeled to allow clear differentiation between its purpose in the workflow. Table 1: Overview of data files/data sets. Data Label Description of Contents Format Purpose Data file 1 Simulated raw task data including task IDs, deadlines, urgency levels, and complexity CSV Base dataset for preprocessing Data file 2 Processed and cleaned task dataset (encoded and normalized attributes) CSV Input for DQN model training Data file 3 Regression analysis dataset with productivity outcomes linked to prioritization results CSV Evaluation of relationship between tasks and productivity Data file 4 Contextual event dataset (calendar events, availability, workload constraints) CSV Supplementary data for adaptive scheduling Data file 5 Python scripts used to simulate and preprocess the data PY Reproducibility of data generation process International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 117 13 Limitations and Future Work While this study demonstrates the potential of reinforcement learning for adaptive task prioritization, several limitations should be considered, primarily stemming from our use of simulated data for this initial proof-of-concept. These limitations, however, provide clear directions for subsequent research. Simulated Data and Generalization: The dataset used for training and evaluation was computationally generated. Although this allowed for controlled experimentation and the precise modeling of specific task attributes like urgency and deadlines, it may not capture the full complexity and unpredictability of real-world human task management. Future work will involve validating the model with data collected from a live user study, which will include unpredictable interruptions, nuanced personal preferences, and real-time schedule changes. Scope of Contextual Factors: The current model incorporates a defined set of contextual factors (e.g., deadlines, calendar events). However, other influential variables, such as a user's fluctuating energy levels, the cognitive load of specific tasks, or interpersonal dependencies in collaborative work, were not modeled. A promising direction for future research is to expand the state space of the DQN to include these additional contextual layers, potentially leveraging data from wearable devices or integrated communication platforms. Model Interpretability: As with many deep learning systems, the DQN operates as a "black box," making it difficult for users to understand the rationale behind a specific prioritization decision. This lack of transparency could impact user trust and adoption. To address this, we plan to integrate explainable AI (XAI) techniques, such as LIME or SHAP, to generate post-hoc explanations for the model's recommendations, fostering greater user confidence and collaboration. Longitudinal Adaptation: The current simulation does not model long-term user habit formation or significant shifts in productivity patterns over weeks or months. The model's ability to adapt to such long-term behavioral changes remains an open question. Future longitudinal studies will be essential to develop mechanisms for the model to continuously learn and forget outdated patterns, ensuring its relevance over extended periods of use. Despite these limitations, this work provides a foundational framework and a strong performance baseline for intelligent task prioritization using reinforcement learning. The addressed future work will be crucial in transitioning the model from a simulated prototype to a robust tool for real-world productivity. Related Works Several studies have addressed the growing importance of intelligent systems in task management and decision support. Early efforts focused on the use of artificial intelligence in International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 118 13 educational and organizational contexts, such as Ali (2020), who reviewed applications of AI in teaching and learning, and Gamper and Knapp (2020), who surveyed intelligent computerassisted learning systems. These studies highlighted the role of AI in enhancing structured activities, but they were limited in addressing the complexity of dynamic task prioritization. Subsequent research has explored the integration of deep reinforcement learning to achieve more adaptive and autonomous decision-making. The foundational work of Mnih et al. (2015) demonstrated the effectiveness of DQNs in complex environments, inspiring their adoption in organizational workflows (Eapen & Liu, 2022; Kim & Park, 2023). Specifically, in the domain of personal productivity, Bader & Matthes (2021) demonstrated the feasibility of DRL for task scheduling in intelligent personal assistants, while works such as Liu et al. (2023) have proposed dedicated frameworks for task prioritization, showing notable improvements over rule-based models In addition, hybrid frameworks combining optimization heuristics and neural models have been proposed. Agostinelli et al. (2021) leveraged Q-networks for heuristic learning in search problems, while Shyalika et al. (2020) provided a review of reinforcement learning applications in dynamic task scheduling. These contributions suggest a strong foundation for applying similar techniques in productivity tools. Researchers have also highlighted the importance of scalability and domain-specific applications. Bhattacharya and Chowdhury (2021) examined AI-driven task management systems in enterprise contexts, whereas Liu et al. (2023) and Yang et al. (2023) reviewed AI-enhanced productivity tools more broadly. Their findings support the notion that adaptive prioritization systems can significantly impact both individual and collaborative work environments. Taken together, the existing literature underscores the need for advanced models that combine learning, adaptability, and interpretability. The present study builds upon these foundations by simulating realistic task management scenarios and proposing reinforcement learning–based solutions that can be extended to real-world platforms. Proposed System and Model Architecture The proposed intelligent task management system is designed around a core reinforcement learning agent responsible for prioritization. The system architecture, illustrated in Figure 1, centers on a Deep Q-Network (DQN) that functions as the prioritization engine, following an architecture proven successful in complex decision-making domains (Mnih et al., 2015). The model is formally framed as a Markov Decision Process (MDP). Users interact with a cross-platform application interface (built with React Native) to input and view their tasks. This front-end communicates with a backend server (built with Django) which manages user data in a SQLite database. The DQN model is hosted on this backend. When tasks International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 119 13 are updated, the backend invokes the DQN, which processes the current list and returns an optimized priority order, relayed back to the user's device. Figure. 1: System architecture Deep Q-Network Model for Prioritization At the core of the system is a Deep Q-Network (DQN) that functions as the prioritization engine. The model is framed as a Markov Decision Process (MDP) where the agent learns an optimal policy for selecting the next task to focus on. 1.State Space (s): The state is a feature vector representing the current situation. For each task, this includes normalized features such as: a.Time until deadline b.Urgency level (categorical, encoded numerically) c.Task complexity rating d.Historical average completion time for similar tasks e.Contextual flags (e.g., alignment with free time windows in calendar) 2.Action Space (a): The action is a discrete choice of which task in the current pending list to select for execution next. 3.Reward Function (r): The reward signal is designed to encourage desirable user outcomes. A positive reward is given for completing a task before its deadline. A large negative reward is assigned if a task misses its deadline. Additional small positive rewards can be granted for maintaining a high overall completion rate, encouraging efficient workflow. International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 120 13 4, Network Architecture: The Q-network is a fully connected neural network with three hidden layers (with ReLU activation functions) that maps the state input to Q-values for each possible action. To ensure stable training, we employ experience replay, where past transitions (s, a, r, s') are stored in a buffer and sampled in mini-batches, and a separate target network that is updated periodically. Integration of Regression Analysis for Model Warm-Up To accelerate the DQN's learning process, we utilize linear regression during an initial warm-up phase. A regression model is trained on historical task data to predict task completion time (y) based on factors like task difficulty (x₁), priority (x₂), and time of day (x₃), using the equation Y = β₀ + β₁x₁ + β₂x₂ + β₃x₃. The predicted completion times from this model are used to preprioritize tasks, creating a more informed initial experience replay buffer. This provides the DQN with a better starting policy than random exploration, allowing it to converge faster to an optimal strategy. Furthermore, the insights from regression analysis help fine-tune the reward function's sensitivity to task duration. Results and Discussions The experimental evaluation of the proposed task prioritization framework demonstrated promising outcomes in terms of both predictive performance and practical efficiency. To rigorously evaluate the performance of our proposed DQN model, we compared it against two standard baseline prioritization methods: A.Earliest Deadline First (EDF): This rule-based algorithm always selects the task with the closest deadline. It is a common-sense baseline that is simple to implement and understand. B.Static Eisenhower Matrix (EM): This method categorizes tasks into four quadrants (Urgent/Important, Not Urgent/Important, etc.) based on fixed rules applied to the 'urgency' and 'complexity' features. Tasks in the 'Urgent and Important' quadrant are always prioritized highest. The DQN's internal learning progress is evidenced by the stable growth and convergence of its cumulative reward over training episodes, as shown in Figure 2. This indicates the successful acquisition of a stable prioritization policy. The final, quantitative results of this comparison are presented in Table 2. The data demonstrates that the DQN significantly outperforms the baselines across all metrics. Most notably, it attained a Deadline Adherence Rate (DAR) of 93.5%, substantially higher than EDF (81.2%) and EM (70.1%). While EDF is designed for deadlines, it fails to account for task complexity and context, a nuance our DQN successfully learns. This leads to a more streamlined workflow, as confirmed by the DQN's lower Average Task Completion Time (4.3 minutes vs. 5.1 and 5.8 for the baselines). International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 121 13 Furthermore, on core classification metrics, the DQN achieved an accuracy of 92.3% and an F1score of 90.0%, compared to 79.5% for EDF and 63.4% for EM. This confirms that the DQN's learning-based approach is fundamentally more effective at correctly identifying truly highpriority tasks than rigid, rule-based systems. The performance of these baselines against our DQN model across key metrics is summarized in Table 2 Figure 2: Learning Curves (Reward and Loss Convergence) Table 2: Comparative performance of the proposed DQN model against baseline prioritization methods. Metric Proposed DQN Model Baseline: Earliest Deadline First (EDF) Accuracy 92.3% 75.1% Precision 89.7% 72.5% Recall 90.4% 88.2% F1 score 90.0% 79.5% Deadline Adherence Rate (DAR) 93.5% 81.2% Average Task Completion Time (ATCT) 4.3 minutes 5.1 minutes