source: papers/FDL2012/abstraction_refinement.tex @ 52

\subsection{Generalities}

We suppose that our concrete model is a composition of several components and each component has been previously verified. Hence, we have a set of verified properties for each component of the concrete model. The main idea of this technique is that we would like to make use of these properties to generate a better abstract model. Properties of the components that appear to be related to the global property to be verified, $\phi$ are selected to generate the abstract model $\widehat{M}_i$. This method is particularly interesting as it gives a possibility to converge quicker to an abstract model that is sufficient to satisfy the global property $\phi$.

\subsubsection{Refinement}
The model-checker provides a counterexample when a property failed during model-checking. The counterexample can be \emph{spurious} which means that the path is impossible in the concrete model $M$ or the counterexample is real which implies that $M \nvDash \phi $.When a counterexample is found to be spurious, it means that the current abstract model $\widehat{M}_i$ is too coarse and has to be refined. In this section, we will discuss about the refinement technique based on the integration of more verified properties of the concrete model's components in the abstract model to be generated. Moreover, the refinement step from $\widehat{M}_i$ to $\widehat{M}_{i+1}$ has to be conservative and respects the properties below:

%\medskip

\begin{property}
All $\widehat{M}_i$ generated are upper-approximations of $M$. Furthermore, we guarantee that $\widehat{M}_{i+1} \sqsubseteq \widehat{M}_i$.
\end{property}
%\bigskip
\begin{property}
$\sigma_i$ is a counterexample of $\widehat{M}_i$ and $\sigma_i$ is not a counterexample of $\widehat{M}_{i+1}$.
\end{property}

%\bigskip
%\newpage

\subsubsection{The Counterexample}


The counterexample at a refinement step $i$, $\sigma_i$ is a path in the abstract model $\widehat{M}_i$ which dissatisfy $\phi$.  In the counterexample given by the model-checker, the variables' value in each states are boolean.
%\medskip

\begin{definition}
\textbf{\emph{The counterexample $\sigma_i$ :}} \\
\\
Let $\widehat{M}_i =(\widehat{AP}_i, \widehat{S}_i, \widehat{S}_{0i}, \widehat{L}_i, \widehat{R}_i, \widehat{F}_i)$ and let the length of the counterexample, $|\sigma_i| = n$: $ \sigma_i = \langle s_{\bar{a}i,0}, s_{\bar{a}i,1}, s_{\bar{a}i,2}, ... , s_{\bar{a}i,k},$ $s_{\bar{a}i,k+1}, ... , s_{\bar{a}i,n}\rangle $ with $ \forall k \in [0,n-1], ~s_{\bar{a}i,k} \subseteq s_{i,k}  \in \widehat{S}_i, ~s_{\bar{a}i,0} \subseteq s_{i,0} \in \widehat{S}_{0i}$ and $(s_{i,k}, s_{i,k+1}) \in \widehat{R}_i$. \\
Furthermore, for each state in $\sigma_i$ we have $s_{\bar{a}i,k} = \langle v_{\bar{a}i,k}^1, v_{\bar{a}i,k}^2, ... ,  v_{\bar{a}i,k}^p, ... , v_{\bar{a}i,k}^q \rangle$, $\forall p \in [1,q], ~v_{\bar{a}i,k}^p \in \widehat{V}_{i,k}$ with $\widehat{V}_{i,k} \in 2^q$. \\
\\
(\emph{\underline{Note} :} In AKS $\widehat{M}_i$, the variables are actually 3-valued : $\widehat{V}_{i,k} \in 3^q$. We differenciate the 3-valued variables  $v_{i,k}^p$ from boolean variables with $v_{\bar{a}i,k}^p$.)\\

%\medskip

\end{definition}

%\bigskip

\begin{definition}
\textbf{\emph{Spurious counterexample :}} \\
\\
Let $\sigma_c = \langle s_{c,0}, s_{c,1}, s_{c,2}, ... , s_{c,k}, s_{c,k+1}, ... , s_{c,n}\rangle$ a path of length $n$ in the concrete model $M$ and in each state of $\sigma_c$ we have $s_{c,k} = \langle v_{c,k}^1, v_{c,k}^2, ... ,  v_{c,k}^{p'}, ... , v_{c,k}^{q'} \rangle$ with $\forall p' \in [1,q'], ~v_{i,k}^{p'} \in V_{c,k}$ and $V_{c,k} \in 2^{q'}$.\\

\smallskip

If $\forall k$ we have $\widehat{V}_{i,k} \subseteq V_{c,k}$ and $\forall v_{\bar{a}i,k} \in \widehat{V}_{i,k}, ~s_{i,k}|_{v_{\bar{a}i,k}} = s_{c,k}|_{v_{c,k}} $ then $M \nvDash \phi$ else $\sigma_i$ is \emph{spurious}.

\end{definition}



\subsection{Pre-processing and pertinency ordering of properties}

Before generating an abstract model to verify a global property $\phi$, the verified properties of all the components in the concrete model are ordered according to their pertinency in comparison to a global property $\phi$. In order to do so, the variable dependency of the variables present in global property $\phi$ has to be analysed. After this point, we refer to the variables present in the global property $\phi$ as \emph{primary variables}.

%\bigskip

The ordering of the properties will be based on the variable dependency graph.
The variables in the model are weighted according to their dependency level
\emph{vis-Ã -vis} primary variables and the properties will be weighted according to the sum of the weights of the variables present in it. We have decided to allocate a supplementary weight for variables which are present at the interface of a component whereas variables which do not interfere in the obtention of a primary variable will be weighted 0. Here is how we proceed:


\begin{enumerate}

\item {\emph{Establishment of primary variables' dependency and maximum graph depth}\\
Each primary variable will be examined and their dependency graph is elobarated. The maximum graph depth among the primary variable dependency graphs will be identified and used to calibrate the weight of all the variables related to the global property.
Given the primary variables of $\phi$, $V_{\phi} =  \langle v_{\phi_0}, v_{\phi_1}, ... , v_{\phi_k}, ... , v_{\phi_n} \rangle$ and $G{\_v_{\phi_k}}$ the dependency graph of primary variable $v_{\phi_k}$, we have the maximum graph depth $max_{d} = max(depth(Gv_{\phi_0}), depth(Gv_{\phi_1}), ... , depth(Gv_{\phi_k}), ... ,$\\$ depth(Gv_{\phi_n})) $.

}

\item {\emph{Weight allocation for each variables} \\
Let's suppose $max_d$ is the maximum dependency graph depth calculated and $p$ is the unit weight. We allocate the variable weight as follows:
\begin{itemize}
\item{All the variables at degree $max_d$ of every dependency graph will be allocated the weight of $p$.}
 \\ \hspace*{20mm} $Wv_{max_d} = p$
\item{All the variables at degree $max_d - 1$ of every dependency graph will be allocated the weight of $2Wv_{max_d}$.}
\\ \hspace*{20mm} $Wv_{max_d - 1} = 2Wv_{max_d}$
\item{...}
\item{All the variables at degree $1$ of every dependency graph will be allocated the weight of $2Wv_{2}$.}
 \\ \hspace*{20mm} $Wv_{1} = 2Wv_{2}$
\item{All the variables at degree $0$ (i.e. the primary variables) will be allocated the weight of $10Wv_{1}$.}
 \\ \hspace*{20mm} $Wv_{0} = 10Wv_{1}$
\end{itemize}

We can see here that the primary variables are given a considerable
ponderation due to their pertinency \emph{vis-Ã -vis} global  property. Furthermore, we will allocate a supplementary weight of $3Wv_{1}$ to variables at the interface of a component as they are the variables which assure the connection between the components if there is at least one variable in the dependency graph established in the previous step in the property. All other non-related variables have a weight equals to $0$.
}


\item {\emph{Ordering of the properties} \\
Properties will be ordered according to the sum of the weight of the variables in it. Therefore, given a property $\varphi_i$ which contains $n+1$ variables, $V_{\varphi_i} =  \langle v_{\varphi_{i0}}, v_{\varphi_{i1}}, ... , v_{\varphi_{ik}}, ... , v_{\varphi_{in}} \rangle$, the weight  of $\varphi_i$ , $W_{\varphi_i} = \sum_{k=0}^{n} Wv_{\varphi_{ik}}$ .
After this stage, we will check all the properties with weight $>0$ and allocate a supplementary weight of $3Wv_{1}$ for every variable at the interface present in the propery. After this process, the final weight of a property is obtained and the properties will be ordered in a list with the weight  decreasing (the heaviest first). We will refer to the ordered list of properties related to the global property $\phi$ as $L_\phi$.


}

\end{enumerate}

%\bigskip

\emph{\underline{Example:}}  \\

For example, if a global property $\phi$ consists of 3 variables: $ p, q, r $ where:
\begin{itemize}
\item{$p$ is dependent of $a$ and $b$}
\item{$b$ is dependent of $c$}
\item{$q$ is dependent of $x$}
\item{$r$ is independent}
\end{itemize}

Example with unit weight= 50.
The primary variables: $p$, $q$ and $r$ are weighted $100x10=1000$ each. \\
The secondary level variables : $a$, $b$ and $x$ are weighted $50x2=100$ each. \\
The tertiary level variable $c$ is weighted $50$. \\
The weight of a non-related variable is $0$.

So each verified properties available pertinency will be evaluated by adding the weights of all the variables in it. It is definitely not an exact pertinency calculation of properties but provides a good indicator of their possible impact on the global property.

\bigskip
\begin{figure}[h!]
   \centering
%   \includegraphics[width=1.2\textwidth]{Dependency_graph_weight_PNG}
%     \hspace*{-15mm}
     \includegraphics{Dependency_graph_weight_PNG}
   \caption{\label{DepGraphWeight} Example of weighting}
\end{figure}

%Dans la figure~\ref{étiquette} page~\pageref{étiquette},  



After this pre-processing phase, we will have a list of properties $L_\phi  $ ordered according to their pertinency in comparison to the global property.




\subsection{Initial abstraction generation}

In the initial abstraction generation, all primary variables have to be represented. Therefore the first element(s) in the list where the primary variables are present will be used to generate the initial abstraction, $\widehat{M}_0$ and we will verify the satisfiability of the global property $\phi$ on this abstract model. If the model-checking failed and the counterexample given is found to be spurious, we will then proceed with the refinement process.



\subsection{Abstraction refinement}

The refinement process from $\widehat{M}_i$ to $\widehat{M}_{i+1}$ can be seperated into 2 steps:

\begin{enumerate}

\item {\emph{\underline{Step 1:}} \\

As we would like to ensure the elimination of the counterexample previously found, we filter out properties that don't have an impact on the counterexample $\sigma_i$ thus won't eliminate it. In order to reach this obective, a Kripke Structure of the counterexample $\sigma_i$, $K(\sigma_i)$ is generated. $K(\sigma_i)$ is a succession of states corresponding to the counterexample path which dissatisfies the global property $\phi$.

\bigskip

\begin{definition}
\textbf{\emph{The counterexample $\sigma_i$ Kripke Structure $K(\sigma_i)$ :}} \\
Let a counterexample of length $n$, $ \sigma_i = \langle s_{\bar{a}i,0}, s_{\bar{a}i,1},\\ s_{\bar{a}i,2}, ... , s_{\bar{a}i,k}, s_{\bar{a}i,k+1}, ... , s_{\bar{a}i,n}\rangle $ with $ \forall k \in [0,n-1]$, we have \\
$K(\sigma_i) = (AP_{\sigma_i}, S_{\sigma_i}, S_{0\sigma_i}, L_{\sigma_i}, R_{\sigma_i})$ a 5-tuple consisting of :

\begin{itemize}
\item { $AP_{\sigma_i}$ : a finite set of atomic propositions which corresponds to the variables in the abstract model $\widehat{V}_{i}$ }      
\item { $S_{\sigma_i} = \{s_{\bar{a}i,0}, s_{\bar{a}i,1}, s_{\bar{a}i,2}, ... , s_{\bar{a}i,k}, s_{\bar{a}i,k+1}, ... , s_{\bar{a}i,n}\}$}
\item { $S_{0\sigma_i} = \{s_{\bar{a}i,0}\}$}
\item { $L_{\sigma_i}$ : $S_{\sigma_i} \rightarrow 2^{AP_{\sigma_i}}$ : a labeling function which labels each state with the set of atomic propositions true in that state. }
\item { $R_{\sigma_i}$ = $ (s_{\bar{a}i,k}, s_{\bar{a}i,k+1})$ }
\end{itemize}
\end{definition}

%\bigskip
All the properties available are then model-checked on $K(\sigma_i)$.

If:
\begin{itemize}
\item {\textbf{$K(\sigma_i) \vDash \varphi  \Rightarrow \varphi $ will not eliminate $\sigma_i$}}
\item {\textbf{$K(\sigma_i) \nvDash \varphi  \Rightarrow \varphi $ will eliminate $\sigma_i$}}
\end{itemize}

%\bigskip


\begin{figure}[h!]
   \centering
%   \includegraphics[width=1.2\textwidth]{K_sigma_i_S_PNG}
%     \hspace*{-15mm}
     \includegraphics{K_sigma_i_S_PNG}
   \caption{\label{AKSNegCex} Kripke Structure of counterexample $\sigma_i$, $K(\sigma_i)$}
\end{figure}

%Dans la figure~\ref{étiquette} page~\pageref{étiquette},  

%\bigskip


\begin{figure}[h!]
   \centering

\begin{tikzpicture}[->,>=stealth',shorten >=1.5pt,auto,node distance=1.8cm,
                    thick]
  \tikzstyle{every state}=[fill=none,draw=blue,text=black]

  \node[initial,state] (A)                            {$s_{\bar{a}i,0}$};
  \node[state]           (B) [below of=A]     {$s_{\bar{a}i,1}$};

  \node[state]           (C) [below of=B]        {$s_{\bar{a}i,k}$};

  \node[state]           (D) [below of=C]       {$s_{\bar{a}i,n-1}$};
  \node[state]           (E) [below of=D]       {$s_{\bar{a}i,n}$};

  \path (A) edge              node {} (B)
            (B) edge       node {} (C)
            (C) edge             node {} (D)
            (D) edge             node {} (E);

\end{tikzpicture}

   \caption{\label{AKSNegCex} Kripke Structure of counterexample $\sigma_i$, $K(\sigma_i)$}
\end{figure}


Therefore all properties that are satisfied won't be chosen to be integrated in the next step of refinement. At this stage, we already have a list of potential properties that will definitely eliminate the current counterexample $\sigma_i$ and might converge the abstract model towards a model sufficient to verify the global property $\phi$.

}
%\bigskip

\item {\emph{\underline{Step 2:}} \\

The property at the top of the list (not yet selected and excluding the properties which are satisfied by $K(\sigma_i)$) is selected to be integrated in the generation of $\widehat{M}_{i+1}$.
%\bigskip

}
\end{enumerate}

$\widehat{M}_{i+1}$ is model-checked and the refinement process is repeated until the model satisfies the global property or there is no property left to be integrated in next abstraction.