1: \begin{abstract}

2: Markov chain Monte Carlo methods are a powerful tool for sampling

3: equilibrium configurations in complex systems. One problem these methods

4: often face is slow convergence over large energy barriers. In this

5: work, we propose a novel method which increases convergence in systems

6: composed of many metastable states. This method aims to connect metastable

7: regions directly using generative neural networks in order to propose

8: new configurations in the Markov chain and optimizes the acceptance

9: probability of large jumps between modes in configuration space. We

10: provide a comprehensive theory and demonstrate the method on example

11: systems.

12: 

13: \begin{keywords}

14: efficient MCMC, high dimensional distribution, invertible models, metastable states

15: \end{keywords}

16: \end{abstract}

17: