DeepFlow: A Multi-Objective Deep Reinforcement Learning Framework for Real-Time Pedagogical Adaptation in Serious Games

The integration of Serious Games and Massive Multiplayer Online RolePlaying Games (MMORPGs) into Computer Science curricula offers a promising avenue for enhancing student engagement and fostering Computational Thinking (CT). However, existing educational MMORPGs predominantly rely on static difficulty curves and heuristic-based scaffolding. This rigidity often fails to accommodate the heterogeneous cognitive profiles of learners, leading to violations of Flow Theory conditions, specifically inducing boredom in advanced students or anxiety in novices. This paper introduces DeepFlow, a novel Deep Reinforcement Learning (DRL) agent directly within the game loop of a generic educational MMORPG architecture. The system aims to function as an autonomous ”Digital Tutor,” dynamically tailoring the learning experience to the individual’s Zone of Proximal Development (ZPD). We formulate the pedagogical adaptation problem as a Multi-Objective Markov Decision Process (MOMDP), where the agent optimizes two conflicting objectives: maximizing learning gains (Concept Mastery) and sustaining engagement (Flow State). Utilizing Proximal Policy Optimization (PPO) with a specialized reward function derived from Item Response Theory (IRT) and Cognitive Load Theory, DeepFlow dynamically adjusts quest complexity, hint granularity, and narrative urgency in real-time. A comprehensive simulation-based evaluation involving varied learner profiles demonstrated that DeepFlow significantly reduces projected attrition rates and improves code comprehension metrics compared to static baselines. Furthermore, behavioral analysis reveals that the DRL agent autonomously discovers complex scaffolding strategies, such as ”fading support,” akin to expert human tutors. While learning gains were positive, we observed that the PPO agent struggled to converge for ’Chaotic’ profiles (agents initialized with random noise), often oscillating between hint levels before stabilizing. This suggests the reward function may need further tuning for highly unpredictable student behaviors. Conclusion: These findings suggest that neuro-adaptive agents can effectively bridge the ”ScalabilityAdaptivity” gap in game-based learning, offering a scalable path toward personalized education.