How we built a self-organizing dual-agent AI system that dynamically generates its own syllabus using PPO and Meta-Learning.
The Vision: What if AI Could Teach Itself?
During the 24-hour SKNCOE Hackathon 2025, our team set out to tackle a provocative problem statement: Autonomous Multi-Objective Curriculum Learning Engine.
Standard machine learning models suffer from rigid training distributions or catastrophic forgetting when faced with complex tasks. We built an AI engine that dynamically generates tasks of increasing difficulty, creating an adaptive feedback loop.
Dual-Agent Architecture & Tech Stack
We architected a Teacher-Learner Dual Agent System:
1. Teacher Agent
- Evaluates the Learner Agent's performance in real time.
- Dynamically adjusts difficulty, task parameters, and optimization objectives.
- Utilizes LLMs for qualitative feedback interpretation and task generation.
2. Learner Agent
- Updates policy using Proximal Policy Optimization (PPO) with Generalized Advantage Estimation (GAE).
- Adapts learning rates using Meta-Learning optimizers.
Tech Deep Dive
- AI/ML Core: Python, PyTorch, PPO, Meta-Learning Optimizers.
- API & Gateway: Python FastAPI + Node.js Express.
- Realtime Telemetry: Socket.io broadcasting convergence metrics to a React dashboard.
Results & Impact
- Mastery Rate: 40% faster convergence compared to fixed-curriculum baseline.
- Hallucination Reduction: Reduced agent error rate during task progression.
- Building an AI that literally learns how to learn proved the power of reinforcement learning combined with meta-learning.