extras.tex

\section{Extras}
\begin{frame}[fragile]{Some useful DRAM options}
  \begin{itemize}
    \item \texttt{ip\_mh\_putOutOfBoundsInChain}:  If true, can be wasteful in
      high-dimensional bounded domains.
    \item \texttt{ip\_mh\_doLogitTransform}:  Transforms bounded domain to
      unbounded domain for efficiency in high dimensions.
    \item \texttt{ip\_mh\_dr\_maxNumExtraStages}:  Number of DR stages to do.
      Can improve mixing.  More stages requires more likelihood evaluations.
    \item \texttt{ip\_mh\_dr\_listOfScalesForExtraStages}:  A list of inverse
      scale factors to scale the proposal covariance matrix by in each DR
      stage.
    \item \texttt{ip\_mh\_am\_initialNonAdaptInterval}:  Number of iterations
      to not adapt.
    \item \texttt{ip\_mh\_am\_adaptInterval}:  The frequency at which to adapt
      proposal covariance matrix (after initial non-adapt period is over).
  \end{itemize}
\end{frame}

\begin{frame}[fragile]{Optimisation}
  \begin{itemize}
    \item \Queso\ can optimise, using deterministic optimisation tools in \Gsl,
      to find the MAP estimator before sampling
    \item Better than starting from a user-defined initial sample, or from a
      realisation from the prior
    \item Caveat:  not all problems are easily optimised
    \item If no derivative provided, then finite differences are used
    \item The default tweakable optimisation parameters are hardcoded.  There is
      an open ticket to change this, but it needs work:
      \url{https://github.com/libqueso/queso/pull/310}
    \item To enable optimisation, call \texttt{seedWithMAPEstimator()} before
      calling \texttt{solveWithBayesMetropolisHastings(\ldots)}.
  \end{itemize}
\end{frame}