Add FAIKR3 Bayesian net reasoning

This commit is contained in:
2023-11-03 17:30:33 +01:00
parent 9f46b9d0af
commit c48c4d4d60
13 changed files with 578 additions and 139 deletions

View File

@ -234,142 +234,3 @@ can be computed as the sum of the atomic events where $\phi$ is true:
\section{Bayesian networks}
\begin{description}
\item[Bayes' rule] \marginnote{Bayes' rule}
\[ \prob{a \,\vert\, b} = \frac{\prob{b \,\vert\, a} \prob{a}}{\prob{b}} \]
\item[Bayes' rule and conditional independence]
Given the random variables $\texttt{Cause}$ and\\
$\texttt{Effect}_1, \dots, \texttt{Effect}_n$, with $\texttt{Effect}_i$ independent from each other,
we can compute $\textbf{P}(\texttt{Cause}, \texttt{Effect}_1, \dots, \texttt{Effect}_n)$ as follows:
\[
\textbf{P}(\texttt{Cause}, \texttt{Effect}_1, \dots, \texttt{Effect}_n) =
\left(\prod_i \textbf{P}(\texttt{Effect}_i \,\vert\, \texttt{Cause})\right) \textbf{P}(\texttt{Cause})
\]
The number of parameters is linear.
\begin{example}
Knowing that $\textbf{P} \models (\texttt{Catch} \perp \texttt{Toothache} \vert \texttt{Cavity})$:
\[
\begin{split}
\textbf{P}&(\texttt{Cavity} \,\vert\, \texttt{toothache} \land \texttt{catch}) \\
&= \alpha\textbf{P}(\texttt{toothache} \land \texttt{catch} \,\vert\, \texttt{Cavity})\textbf{P}(\texttt{Cavity}) \\
&= \alpha\textbf{P}(\texttt{toothache} \,\vert\, \texttt{Cavity})
\textbf{P}(\texttt{catch} \,\vert\, \texttt{Cavity})\textbf{P}(\texttt{Cavity}) \\
\end{split}
\]
\end{example}
\item[Bayesian network] \marginnote{Bayesian network}
Graph for conditional independence assertions and a compact specification of full joint distributions.
\begin{itemize}
\item Directed acyclic graph.
\item Nodes represent variables.
\item The conditional distribution of a node is given by its parents
\[ \textbf{P}(X_i \,\vert\, \texttt{parents}(X_i)) \]
In other words, if there is an edge from $A$ to $B$, then $A$ (cause) influences $B$ (effect).
\end{itemize}
\begin{description}
\item[Conditional probability table (CPT)] \marginnote{Conditional probability table (CPT)}
In the case of boolean variables, the conditional distribution of a node can be represented using
a table by considering all the combinations of the parents.
\begin{example}
Given the boolean variables $A$, $B$ and $C$, with $C$ depending on $A$ and $B$, we have that:\\
\begin{minipage}{.48\linewidth}
\centering
\includegraphics[width=0.35\linewidth]{img/_cpt_graph.pdf}
\end{minipage}
\begin{minipage}{.48\linewidth}
\centering
\begin{tabular}{c|c|c|c}
A & B & $\prob{c \vert A, B}$ & $\prob{\lnot c \vert A, B}$ \\
\hline
a & b & $\alpha$ & $1-\alpha$ \\
$\lnot$a & b & $\beta$ & $1-\beta$ \\
a & $\lnot$b & $\gamma$ & $1-\gamma$ \\
$\lnot$a & $\lnot$b & $\delta$ & $1-\delta$ \\
\end{tabular}
\end{minipage}
\end{example}
\end{description}
\item[Reasoning patterns] \marginnote{Reasoning patterns}
Given a Bayesian network, the following reasoning patterns can be used:
\begin{descriptionlist}
\item[Causal] \marginnote{Causal reasoning}
To make a prediction. From the cause, derive the effect.
\begin{example}
Knowing $\texttt{Intelligence}$, it is possible to make a prediction of $\texttt{Letter}$.
\begin{center}
\includegraphics[width=0.5\linewidth]{img/_causal_example.pdf}
\end{center}
\end{example}
\item[Evidential] \marginnote{Evidential reasoning}
To find an explanation. From the effect, derive the cause.
\begin{example}
Knowing $\texttt{Grade}$, it is possible to explain it by estimating\\$\texttt{Intelligence}$.
\begin{center}
\includegraphics[width=0.65\linewidth]{img/_evidential_example.pdf}
\end{center}
\end{example}
\item[Explain away] \marginnote{Explain away reasoning}
Observation obtained "passing through" other observations.
\begin{example}
Knowing $\texttt{Difficulty}$ and $\texttt{Grade}$,
it is possible to estimate \\$\texttt{Intelligence}$.
Note that if $\texttt{Grade}$ was not known,
$\texttt{Difficulty}$ and $\texttt{Intelligence}$ would be independent.
\begin{center}
\includegraphics[width=0.70\linewidth]{img/_explainaway_example.pdf}
\end{center}
\end{example}
\end{descriptionlist}
\item[Global semantics] \marginnote{Global semantics}
Given a Bayesian network, the full joint distribution can be defined as
the product of the local conditional distributions:
\[ \prob{x_1, \dots, x_n} = \prod_{i=1}^{n} \prob{x_i \,\vert\, \texttt{parents}(X_i)} \]
\begin{example}
Given the following Bayesian network:
\begin{minipage}{.3\linewidth}
\centering
\includegraphics[width=0.7\linewidth]{img/_global_semantics_example.pdf}
\end{minipage}
\begin{minipage}{.6\linewidth}
\[
\begin{split}
&\prob{j \land m \land a \land \lnot b \land \lnot e} \\
&= \prob{\lnot b} \prob{\lnot e} \prob{a \,\vert\, \lnot b, \lnot e}
\prob{j \,\vert\, a} \prob{m \,\vert\, a}
\end{split}
\]
\end{minipage}
\end{example}
\item[Independence] \marginnote{Bayesian network independence}
Intuitively, an effect is independent from a cause,
if there is another cause in the middle whose value is already known.
\begin{example}
\phantom{}
\begin{minipage}{.3\linewidth}
\centering
\includegraphics[width=0.75\linewidth]{img/_independence_example.pdf}
\end{minipage}
\begin{minipage}{.6\linewidth}
\[ \textbf{P} \models (\texttt{L} \perp \texttt{D}, \texttt{I}, \texttt{S} \,\vert\, \texttt{G}) \]
\[ \textbf{P} \models (\texttt{S} \perp \texttt{L} \,\vert\, \texttt{G}) \]
\[ \textbf{P} \models (\texttt{S} \perp \texttt{D}) \text{ but }
\textbf{P} \models (\texttt{S} \,\cancel{\perp}\, \texttt{D} \,\vert\, \texttt{G}) \text{ (explain away)} \]
\end{minipage}
\end{example}
\end{description}