Polish talk a bit
This commit is contained in:
parent
ac2e9111a6
commit
1a86506447
|
@ -31,3 +31,14 @@
|
|||
journaltitle = {Siemens AG, CT IC 4, Technical Report},
|
||||
title = {The Wet Game of Chicken}
|
||||
}
|
||||
|
||||
@article{kaiser_interpretable_2019,
|
||||
abstract = {In this paper, we present a Bayesian view on model-based reinforcement learning. We use expert knowledge to impose structure on the transition model and present an efficient learning scheme based on variational inference. This scheme is applied to a heteroskedastic and bimodal benchmark problem on which we compare our results to NFQ and show how our approach yields human-interpretable insight about the underlying dynamics while also increasing data-efficiency.},
|
||||
author = {Kaiser, Markus and Otte, Clemens and Runkler, Thomas and Ek, Carl Henrik},
|
||||
date = {2019},
|
||||
journaltitle = {Computational Intelligence and Machine Learning},
|
||||
langid = {english},
|
||||
pages = {6},
|
||||
title = {Interpretable {{Dynamics Models}} for {{Data}}-{{Efficient Reinforcement Learning}}},
|
||||
volume = {ESANN 2019 proceedings}
|
||||
}
|
||||
|
|
|
@ -105,7 +105,7 @@
|
|||
}
|
||||
|
||||
|
||||
\includeonlyframes{policy}
|
||||
% \includeonlyframes{summary}
|
||||
|
||||
|
||||
\begin{document}
|
||||
|
@ -136,7 +136,7 @@
|
|||
\end{frame}
|
||||
|
||||
|
||||
\begin{frame}[label=dynamics_posterior]{Multimodal System Dynamics}
|
||||
\begin{frame}[label=dynamics_posterior]{Multimodal System Dynamics\footcite{kaiser_interpretable_2019}}
|
||||
\centering
|
||||
\includestandalonewithpath{figures/dynamics_posterior_cut}
|
||||
\end{frame}
|
||||
|
@ -146,17 +146,21 @@
|
|||
\centering
|
||||
\begin{columns}[T]
|
||||
\begin{column}{.45\textwidth}
|
||||
\centering
|
||||
\includestandalone{figures/policy_quiver}
|
||||
\begin{align}
|
||||
\Fun*{R}{x, y} = x
|
||||
\end{align}
|
||||
22\% drop rate
|
||||
\end{column}
|
||||
\begin{column}{.45\textwidth}
|
||||
\centering
|
||||
\uncover<2>{
|
||||
\includestandalone{figures/conservative_policy_quiver}
|
||||
\begin{align}
|
||||
\Fun*{R^\prime}{x, y} = \Fun*{R}{x, y} - 5 \cdot \Prob{\text{drop} \given x, y}
|
||||
\end{align}
|
||||
19\% drop rate
|
||||
}
|
||||
\end{column}
|
||||
\end{columns}
|
||||
|
@ -165,30 +169,30 @@
|
|||
|
||||
\begin{frame}[label=summary]{Data Association with Gaussian Processes}
|
||||
\medskip
|
||||
\begin{block}{Interaction with domain experts}
|
||||
\begin{block}{Model for multimodal data}
|
||||
\vspace{1ex}
|
||||
\begin{columns}[c]
|
||||
\begin{column}[c]{.5\textwidth}
|
||||
\begin{itemize}
|
||||
\item Incomplete system knowledge
|
||||
\item Hierarchical priors
|
||||
\item Interpretable sub-models
|
||||
\item Separate models per mode
|
||||
\item Predictive Associations
|
||||
\item Scalable inference
|
||||
\end{itemize}
|
||||
\end{column}
|
||||
\begin{column}[c]{.475\textwidth}
|
||||
\centering
|
||||
\includestandalonewithpath{figures/graphical_model_mdgp_tiny}
|
||||
\includestandalonewithpath{figures/graphical_model_dagp_tiny}
|
||||
\end{column}
|
||||
\end{columns}
|
||||
\end{block}
|
||||
%
|
||||
\begin{block}{Trustworthy decision making}
|
||||
\begin{block}{Informed decision making}
|
||||
\begin{columns}[c]
|
||||
\begin{column}[c]{.5\textwidth}
|
||||
\begin{itemize}
|
||||
\item Uncertainty due to incomplete data
|
||||
\item Hierarchical priors
|
||||
\item Interpretable sub-models
|
||||
\item Stochastic systems
|
||||
\item Robust and efficient inference
|
||||
\end{itemize}
|
||||
\end{column}
|
||||
\begin{column}[c]{.475\textwidth}
|
||||
|
|
|
@ -12,24 +12,16 @@
|
|||
\def\datapath{\figurepath/data}
|
||||
|
||||
\tikzstyle{tiny random variable} = [random variable, inner sep=0pt, minimum size=13pt, font=\tiny]
|
||||
\node[tiny random variable, observed, xshift=-2ex] (Xsn) at (1, 0.25) {$\rv{s}_t$};
|
||||
\node[tiny random variable, observed, xshift=2ex] (Xan) at (1, 0.25) {$\rv{a}_t$};
|
||||
\node[tiny random variable, observed] (Xn) at (1, 0) {$\rv{x_n}$};
|
||||
\node[tiny random variable, observed] (yn) at (1, -3) {$\rv{y_n}$};
|
||||
|
||||
\begin{scope}[on background layer]
|
||||
\node[plate, opacity=0.3, fit=(Xsn)(Xan)] (Xn) {};
|
||||
\end{scope}
|
||||
\node[tiny random variable, observed] (yn) at (.8, -3) {$\rv{s}_{t+1}$};
|
||||
\node[tiny random variable, latent] (fnm) at (0, -1) {$\rv{f_n^{\pix{k}}}$};
|
||||
\node[tiny random variable, latent] (ynm) at (0, -2) {$\rv{y_n^{\pix{k}}}$};
|
||||
|
||||
\node[tiny random variable, latent] (fnm) at (0, -1) {$\rv{f}_t^{\pix{k}}$};
|
||||
\node[tiny random variable, latent] (sigmanm) at (1, -1) {$\rv{\sigma}_t^{\pix{k}}$};
|
||||
\node[tiny random variable, latent] (ynm) at (0, -2) {$\rv{s}_{t+1}^{\pix{k}}$};
|
||||
|
||||
\node[tiny random variable, latent] (alphanm) at (2, -1) {$\rv{\lambda}_t^{\pix{k}}$};
|
||||
\node[tiny random variable, latent] (an) at (2, -3) {$\rv{l}_t$};
|
||||
\node[tiny random variable, latent] (alphanm) at (2, -1) {$\rv{\alpha_n^{\pix{k}}}$};
|
||||
\node[tiny random variable, latent] (an) at (2, -3) {$\rv{a_n}$};
|
||||
|
||||
\draw[edge, directed] (Xn) -| (fnm);
|
||||
\draw[edge, directed] (Xn) -- (sigmanm);
|
||||
\draw[edge, directed] (sigmanm) -- (ynm);
|
||||
\draw[edge, directed] (Xn) -| (alphanm);
|
||||
\draw[edge, directed] (fnm) -- (ynm);
|
||||
\draw[edge, directed] (alphanm) -- (an);
|
||||
|
@ -40,8 +32,8 @@
|
|||
|
||||
\begin{scope}[on background layer]
|
||||
\node[
|
||||
yshift=3pt,
|
||||
plate, inner sep=5pt,
|
||||
% yshift=3pt,
|
||||
plate, inner sep=3pt,
|
||||
opacity=0.3,
|
||||
fit=(fnm)(alphanm)(ynm),
|
||||
label={[font=\small, opacity=0.3, anchor=south east]south east:K}
|
|
@ -27,7 +27,7 @@
|
|||
point meta={-abs(\thisrow{aR})},
|
||||
] table[
|
||||
ignore chars={\#}, col sep=space, x=Y, y=X,
|
||||
] {\datapath/policy_quiver.dat};
|
||||
] {\datapath/conservative_policy_quiver.dat};
|
||||
|
||||
\addplot[
|
||||
quiver={%
|
||||
|
@ -41,7 +41,7 @@
|
|||
},
|
||||
] table[
|
||||
ignore chars={\#}, col sep=space, x=Y, y=X,
|
||||
] {\datapath/policy_quiver.dat};
|
||||
] {\datapath/conservative_policy_quiver.dat};
|
||||
\end{axis}
|
||||
|
||||
% NOTE(mrksr): Hacky fix for tikzpicture in block
|
||||
|
|
|
@ -53,7 +53,7 @@
|
|||
\node[text width=7cm] at (2.25, 0.5) {
|
||||
\begin{description}
|
||||
\item[Dynamics] Agent in a flowing river
|
||||
\item[Goal] Get close to waterfall
|
||||
\item[Goal] Get close to the waterfall
|
||||
\item[State] $(x, y)$-position in $\R^2$
|
||||
\item[Action] $(x, y)$-movement in $\R^2$
|
||||
\end{description}
|
||||
|
|
Loading…
Reference in New Issue