A deep reinforcement learning agent for geometry online tutoring

In this paper, we apply deep reinforcement learning (DRL) for geometry reasoning and develop Dragon to facilitate online tutoring. Its success is contingent on a flexible data model to capture diverse concepts and heterogeneous relations, as well as an effective DRL agent to generate near-optimal and human-readable solutions. We use proximal policy optimization (PPO) as the backbone DRL architecture, customized with effective state representation and integrated with a bunch of optimization tricks including attention mechanism, action mask, data augmentation and curriculum learning. In our experimental study, we craft so far the largest scale dataset with geometry problems and a knowledge base with 46 theorems. We implement various heuristic algorithms and DRL models as baselines for performance comparison. The results show that our agent achieves near-optimal solution and is superior over multiple competitive baselines. To benefit the community, we opensource the dataset and implementation at https://github.com/AIEdu-xzy/geometry-solver .