Mask the reinforcement learning policy's action space to a zone-of-proximal-development difficulty band (success probability 0.4–0.8)¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The adaptive training component treats content sequencing as a sequential decision problem solved with Proximal Policy Optimization. To formalize the zone of proximal development, content whose estimated success probability falls outside 0.4–0.8 is excluded: "Content falling outside this range is masked from the policy’s action space during training and inference, ensuring that the RL agent can only recommend activities within a pedagogically meaningful difficulty band." The reward combines diagnostic score improvement, task completion rate, and time-on-task efficiency.

Design Implications¶

Context¶

Requirements¶

• A trained diagnostic model supplying learner ability states, and empirically calibrated difficulty bounds (0.4 and 0.8) validated on a held-out set

Constraints¶

• Bounds were calibrated by maximizing normalized learning gain on a held-out validation set; sensitivity analysis showed stable performance only across moderate variations of about ±0.05

Target Learners¶

• University English learners across proficiency levels

Target Learning Goals¶

• Durable oral proficiency gains through appropriately challenged, individualized practice sequencing

Related Strategies¶

• Scaffolded Difficulty Progression
• Scaffolded Questioning

Examples¶

-

Key Sources¶

• Chai Rui. (2026). Deep learning-based intelligent diagnosis and adaptive training system for university english oral proficiency. Scientific Reports. https://doi.org/10.1038/s41598-026-51608-6