Welcome to DynGPT’s documentation!

PyPI PyPI - Python Version

Overview of DynGPT framework

DynGPT is a GPT-driven framework designed to solve generalized multi-dimensional state transition networks (STNs) and learn stochastic dynamics from static observed data.

Learning stochastic dynamic from static observed data

DynGPT consists of two core modules: DynGPT-Solver and DynGPT-Inferrer. DynGPT-Solver effectively captures the complex dynamical properties of stationary distributions in intractable STNs across parameter spaces. Leveraging the trained DynGPT-Solver and neural approximate Bayesian computation (NeuralABC), DynGPT-Inferrer efficiently and accurately estimates the posterior distributions of STN parameters from multi-dimensional static observed data.

DynGPT-Solver: solving the stationary distribution of the state transition networks

Solving the steady-state distribution of complex STNs remains challenging, as analytical solutions are often intractable, while numerical approaches typically require significant computational resources or compromise accuracy for efficiency. To address this, DynGPT-Solver employs an autoregressive transformer-based architecture to efficiently solve the joint stationary distribution of multi-dimensional STNs.

Learning stochastic dynamic from static observed data

DynGPT-Inferrer: learning stochastic dynamic from static observed data

Understanding how stochasticity enhances robustness and flexibility is a key focus in systems dynamics research, requiring the modeling and inference of underlying stochastic mechanisms from observed data. DynGPT-Inferrer efficiently estimates the Bayesian posterior distributions of STN parameters by utilizing automatic differentiation and NeuralABC.

Learning stochastic dynamic from static observed data

Reference

DynGPT: a generative AI framework for learning stochastic dynamics from static data through state transition networks

Contents