Abstract
Modern artificial intelligence (AI) methods have been used to solve problems that many humans struggle to solve. This opens up new opportunities for knowledge discovery and education. We demonstrate ALLURE, an educational AI system for learning to solve the Rubik’s cube that is designed to help students improve their problem-solving skills. ALLURE can both find and explain its own strategies for solving the Rubik’s cube as well as build on user-provided strategies. Collaboration between AI and user happens using visual and natural language modalities.
Introduction: The Challenge of Explainable AI in Education
While modern AI solvers like DeepCubeA can solve a Rubik’s Cube with superhuman efficiency, they typically operate as "black boxes." They provide a sequence of moves to reach a solved state but fail to explain the underlying logic or strategy. For a human learner, especially a child, a list of twenty random rotations is pedagogically useless. Humans learn through a hierarchy of subgoals—intermediate milestones that break a complex problem into manageable parts. The ALLURE framework (Automatic Learning and Logic-based Understanding for Rubik’s cube Education) addresses this gap by transforming raw AI move sequences into human-understandable strategic concepts.
Methodology: Integrating RL, ILP, and NLP
The ALLURE system integrates three core AI technologies to create a collaborative learning environment:
- Deep Reinforcement Learning (DeepCubeA): The core solver uses DeepCubeA combined with Hindsight Experience Replay (HER). This allows the model to be trained specifically to achieve subgoals (partial configurations) rather than just the final solved state.
- Inductive Logic Programming (ILP): To bridge the gap between raw rotations and human language, the system utilizes the Popper ILP engine. ILP extracts relational predicates and logical rules from the AI’s move sequences. This process turns a sequence like "Front-Clockwise, Right-Counterclockwise" into a higher-level strategic concept.
- Natural Language Processing and Affective Computing: Using the Rasa framework, ALLURE features an interactive chatbot. This bot translates induced logic into natural language hints and monitors user sentiment. By tracking frustration or engagement, the AI can adjust its pedagogical approach, offering encouragement or simplifying its explanations as needed.
Subgoals as a Pedagogical Interface
The most critical insight of the ALLURE framework is that subgoals serve as the "common tongue" between AI search and human cognition. By forcing the AI to solve for intermediate milestones, its internal search process becomes alignable with traditional human teaching methods.
For example, the "white cross" on the top face is a standard first step in human solving methods. ALLURE can identify this subgoal, find the optimal path to it, and then explain the logic of that path using symbolic rules that are inherently more translatable into human language than standard saliency maps.
Key Results: Interactive Visual Guidance
The current ALLURE prototype successfully demonstrates strategic decomposition and automated explanation. The system's multi-modal user interface combines 3D visualizations with real-time animations and voice/text guidance.
As shown in the visual sequence above, the system guides the user through necessary rotations (e.g., D', F', R) while explaining the effect of each move on the cube's state. This active learning approach ensures the student understands why a move is made, not just what move to make.
Limitations and Future Work
While the prototype is functional, several areas for future development remain. The authors plan to conduct empirical user studies with children to measure the actual improvement in problem-solving skills compared to traditional video tutorials. Furthermore, they aim to implement bidirectional collaboration, allowing users to propose their own subgoals or algorithms for the AI to analyze and optimize. Finally, the framework's combination of RL, ILP, and NLP holds potential for generalization to other complex planning and search problems beyond the Rubik’s Cube.
Summary updated