X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/desynchronisation-controle.git/blobdiff_plain/bc18d50297e1b6798015a28bc91eee4f7f3752ca..1216fb04e58a581d9a4d2e77355aaf47bf989447:/IWCMC14/HLG.tex diff --git a/IWCMC14/HLG.tex b/IWCMC14/HLG.tex index 9833667..ff12b04 100644 --- a/IWCMC14/HLG.tex +++ b/IWCMC14/HLG.tex @@ -1,56 +1,45 @@ -Let us first the basic recalls of the~\cite{HLG09} article. - - -The precise the context of video sensor network as represeted for instance -in figure~\ref{fig:sn}. - \begin{figure} \begin{center} -\includegraphics[scale=0.5]{reseau.png} +\includegraphics[scale=0.3]{reseau.png} + +\begin{scriptsize} +An example of a sensor network ofsize 10. All nodes are video sensor +except the 5 and the 9 one which is the sink. +\JFC{reprendre la figure, trouver un autre titre} +\end{scriptsize} + \caption{SN with 10 sensor}\label{fig:sn}. \end{center} \end{figure} - -Let us give a formalisation of such a wideo network sensor. -We start with the flow formalising: - -The video sensor network is represented as a strongly -connected oriented labelled graph. +Let us first recall the basics of the~\cite{HLG09} article. +The video sensor network is memorized as a connected non oriented +oriented labelled graph. In this one, -the nodes, in a set $N$ are sensors, links, or the sink. -Furtheremore, there is an edge from $i$ to $j$ if $i$ can -send a mesage to $j$. The set of all edges is further denoted as -$L$ . -This boolean information is stored as a +the nodes, in a set $N$, are sensors, links, or the sink. +Furthermore, there is an edge from $i$ to $j$ if $i$ can +send a message to $j$. The set of all edges is further denoted as +$L$. +Figure~\ref{fig:sn} gives an example of such a network. + +This link information is stored as a matrix $A=(a_{il})_{i \in N, l \in L}$, where -$a_{il} = -\left\{ - \begin{array}{rl} - 1 & \textrm{if $l$ starts with $i$ } \\ - -1 & \textrm{si $l$ ends width $i$ } \\ - 0 & \textrm{otherwise} - \end{array} - \right.$. +$a_{il}$ is 1 if $l$ starts with $i$, is -1 if $l$ ends width $i$ +and 0 otherwise. Let $V \subset N $ be the set of the video sensors of $N$. -Let thus $R_h$, $R_h \geq 0$ be the encoding rate of video sensor $h$, $h \in V$. -Let $\eta_{hi}$ be the production rate of the $i$ node, for the $h$ session. More precisely, we have - $$ -\eta_{hi} = -\left\{ - \begin{array}{rl} - R_h & \textrm{si $i$ est $h$} \\ - -R_h & \textrm{si $i$ est le puits} \\ - 0 & \textrm{sinon} - \end{array} - \right.$$ +Let thus $R_h$, $R_h \geq 0$, +be the encoding rate of video sensor $h$, $h \in V$. +Let $\eta_{hi}$ be the production rate of the node $i$, +for the session initiated by $h$. More precisely, we have +$ \eta_{hi}$ is equal to $ R_h$ if $i$ is $h$, +is equal to $-R_h$ if $i$ is the sink, and $0$ otherwise. We are then left to focus on the flows in this network. Let $x_{hl}$, $x_{hl}\geq 0$, be the flow inside the edge $l$ that -issued from the $h$ sesssion and +issued from the $h$ session and let $y_l = \sum_{h \in V}x_{hl} $ the sum of all the flows inside $l$. Thus, what is produced inside the $i^{th}$ sensor for session $h$ is $ \eta_{hi} = \sum_{l \in L }a_{il}x_{hl} $. @@ -62,7 +51,7 @@ The distortion is bounded $\sigma^2 e^{-\gamma . R_h.P_{sh}^{}2/3} \leq D_h$. The initial energy of the $i$ node is $B_i$. -The overall consumed powed of the $i$ node is +The overall consumed power of the $i$ node is $P_{si}+ P_{ti} + P_{ri}= P_{si}+ \sum_{l \in L}a_{il}^{+}.c^s_l.y_l + \sum_{l \in L} a_{il}^{-}.c^r.y_l \leq q.B_i. @@ -81,24 +70,135 @@ The objective is thus to find $R$, $x$, $P_s$ which minimize \item $P_{sh} > 0,\forall h \in V$ \end{enumerate} - -To achieve a local optimisation, the problem is translated into an -equivalent one: - -\begin{itemize} -The objective is thus to find $R$, $x$, $P_s$ which minimize +To achieve this optimizing goal +a local optimisation, the problem is translated into an +equivalent one: find $R$, $x$, $P_s$ which minimize $\sum_{i \in N }q_i^2$ with the same set of constraints, but item \ref{itm:q}, which is replaced by: +$$ +\begin{array}{l} +P_{si}+ \sum_{l \in L}a_{il}^{+}.c^s_l.\left( \sum_{h \in V}x_{hl} \right) \\ +\qquad + + \sum_{l \in L} a_{il}^{-}.c^r.\left( \sum_{h \in V}x_{hl} \right) \leq q.B_i, \forall i \in N +\end{array} +$$ + + +They thus replace the objective of reducing +$\sum_{i \in N }q_i^2$ +by the objective of reducing +\begin{equation} +\sum_{i \in N }q_i^2 + +\sum_{h \in V, l \in L } \delta.x_{hl}^2 ++ \sum_{h \in V }\delta.R_{h}^2 +\label{eq:obj2} +\end{equation} +where $\delta$ is a regularisation factor. +This indeed introduces quadratic fonctions on variables $x_{hl}$ and +$R_{h}$ and makes some of the functions strictly convex. + +The authors then apply a classical dual based approach with Lagrange multiplier +to solve such a problem~\cite{}. +They first introduce dual variables +$u_{hi}$, $v_{h}$, $\lambda_{i}$, and $w_l$ for any +$h \in V$,$ i \in N$, and $l \in L$. + +\begin{equation} +\begin{array}{l} +L(R,x,P_{s},q,u,v,\lambda,w)=\\ +\sum_{i \in N} q_i^2 + q_i. \left( +\sum_{l \in L } a_{il}w_l^{(k)}- +\lambda_iB_i +\right)\\ ++ \sum_{h \in V} +v_h.\dfrac{\ln(\sigma^2/D_h)}{\gamma P_{sh} ^{2/3}} + \lambda_h P_{sh}\\ ++ \sum_{h \in V} \sum_{l\in L} +\left( +\delta.x_{hl}^2 \right.\\ +\qquad \qquad + x_{hl}. +\sum_{i \in N} \left( +\lambda_{i}.(c^s_l.a_{il}^{+} + +c^r. a_{il}^{-} ) \right.\\ +\qquad \qquad\qquad \qquad + +\left.\left. u_{hi} a_{il} +\right) +\right)\\ + + \sum_{h \in V} +\delta R_{h}^2 +-v_h.R_{h} - \sum_{i \in N} u_{hi}\eta_{hi} +\end{array} +\end{equation} + +The proposed algorithm iteratively computes the following variables +untill the variation of the dual function is less than a given threshold. +\begin{enumerate} +\item $ u_{hi}^{(k+1)} = u_{hi}^{(k)} - \theta^{(k)}. \left( + \eta_{hi}^{(k)} - \sum_{l \in L }a_{il}x_{hl}^{(k)} \right) $ +\item +$v_{h}^{(k+1)}= \max\left\{0,v_{h}^{(k)} - \theta^{(k)}.\left( R_h^{(k)} - \dfrac{\ln(\sigma^2/D_h)}{\gamma.(P_{sh}^{(k)})^{2/3}} \right)\right\}$ +\item + $\begin{array}{l} + \lambda_{i}^{(k+1)} = \max\left\{0, \lambda_{i}^{(k)} - \theta^{(k)}.\left( + q^{(k)}.B_i \right. \left.\\ + \qquad\qquad\qquad -\sum_{l \in L}a_{il}^{+}.c^s_l.\left( \sum_{h \in V}x_{hl}^{(k)} \right) \\ + \qquad\qquad\qquad - \left.\left. \sum_{l \in L} a_{il}^{-}.c^r.\left( \sum_{h \in V}x_{hl}^{(k)} \right) - P_{si}^{(k)} \right) \right\} +\end{array} +$ -$$P_{si}+ \sum_{l \in L}a_{il}^{+}.c^s_l.\left( \sum_{h \in V}x_{hl} \right) + -\sum_{l \in L} a_{il}^{-}.c^r.\left( \sum_{h \in V}x_{hl} \right) \leq q.B_i, \forall i \in N$$ - -\JFC{Vérifier l'inéquation précédente} +\item +$w_l^{(k+1)} = w_l^{(k+1)} + \theta^{(k)}. \left( \sum_{i \in N} a_{il}.q_i^{(k)} \right)$ -The authors then aplly a dual based approach with Lagrange multiplier -to solve such a problem. +\item +$\theta^{(k)} = \omega / t^{1/2}$ + \item +$q_i^{(k)} = \arg\min_{q_i>0} +\left( +q_i^2 + q_i. +\left( +\sum_{l \in L } a_{il}w_l^{(k)}- +\lambda_i^{(k)}B_i +\right) +\right)$ +\item \label{item:psh} +$ +P_{sh}^{(k)} += +\arg \min_{p > 0} +\left( +v_h^{(k)}.\dfrac{\ln(\sigma^2/D_h)}{\gamma p ^{2/3}} + \lambda_h^{(k)}p +\right) +$ -\inputFrameb{Formulation simplifiée}{formalisationsimplifiee} \ No newline at end of file +\item +$ +R_h^{(k)} += +\arg \min_{r \geq 0 } +\left( +\delta r^2 +-v_h^{(k)}.r - \sum_{i \in N} u_{hi}^{(k)} \eta_{hi} +\right) +$ +\item +$ +x_{hl}^{(k)} = +\begin{array}{l} +\arg \min_{x \geq 0} +\left( +\delta.x^2 \right.\\ +\qquad \qquad + x. +\sum_{i \in N} \left( +\lambda_{i}^{(k)}.(c^s_l.a_{il}^{+} + +c^r. a_{il}^{-} ) \right.\\ +\qquad \qquad\qquad \qquad + +\left.\left. u_{hi}^{(k)} a_{il} +\right) +\right) +\end{array} +$ +\end{enumerate} +where the first four elements are dual variable and the last four ones are +primal ones. \ No newline at end of file