rl

What problem does it solve?

This Skill captures a reproducible, self-contained dual-agent reinforcement learning workflow that demonstrates epistemic reward shaping, enabling researchers to study how a teacher network can shape a student's learning via task-specific reward maps and per-task delta signals, improving generalization to unseen environments.

Core Features & Use Cases

• Dual PPO Student networks with separate optimizers and memory buffers.
• Teacher network that outputs an 8x8 reward map conditioned on an epistemic state.
• Deterministic, notebook-based training without external RL libraries.
• Per-task delta-based teacher reward, per-task success tracking, and a curriculum of 10 training tasks plus 3 unseen test tasks.
• ONNX export for frontend visualization and JSON/episode logs for analysis.
• Generalization evaluation on unseen gridworld layouts to quantify transfer.

Quick Start

Open the epistemic_rl.ipynb notebook and run the training loop for the teacher-guided or baseline random conditions to reproduce PPOStudent and TeacherAgent interactions and export ONNX models.