Dependence of convergence on random seed

Usually I will analysis the cases where it does not work:

  • lots of divergence. solution: more informative prior
  • some chain is a flat line. solution: likely initial start point problem, manually supply start
  • bad mixing (random walk). solution: this is a tricky one… likely the mass matrix is adapted to a local geometric, manually start point might help. otherwise, you might want to try smarter mass matrix adaptation like the one in https://github.com/dfm/exoplanet
2 Likes