Active Learning for Dynamical Systems using Bayesian Neural Networks

This project explores data-efficient learning of unknown dynamical systems using active learning with Bayesian Neural Networks (BNNs). When collecting real-world data is costly, actively selecting informative samples can significantly improve sample efficiency.

Key Features

• Bayesian Neural Network Model: Learns the system dynamics while quantifying uncertainty.
  • Monte-Carlo Dropout
  • Laplace Approximation
• Active Learning Strategies:
  • Random Sampling Shooting with Model Predictive Control
  • Soft Actor Critic (SAC) policy optimization
• Feature Expansion: Ensures uncertainty estimates remain invariant to input norm.
• Custom Environments:
  • Mass-Spring-Damper System
  • 2D Reacher Task

Evaluation

We evaluate performance based on:

• One-step and multi-step prediction error
• State space exploration efficiency

Results

Our methods show superior sample efficiency compared to random exploration, thanks to uncertainty-guided data collection.

Code and Experiments

You can find the full implementation and experiments in this repository.

TUM-ADLR-10/

├── environments/ # Custom environments

├── experiments/ # Scripts, results and saved data for experiments

├── metrics/ # Evaluation metrics

├── models/ # Neural network model definitions

├── sampling_methods/ # Active learning exploration algorithms

├── utils/ # Utility functions

└── weights/ # Trained model weights

Poster and Paper

This repository is based on our course project for TUM ADLR course. For presentation poster and report paper, see folder presentation/ .