totot

[desynchronisation-controle.git] / IWCMC14 / convexity.tex

diff --git a/IWCMC14/convexity.tex b/IWCMC14/convexity.tex
index 172ca3734b5da9936565f3f0e0f360c1e33715be..d88481cf4a58773f0ded55ea3b816b4766826232 100644 (file)
--- a/IWCMC14/convexity.tex
+++ b/IWCMC14/convexity.tex
@@ -12,6 +12,10 @@ The function inside the $\arg \min$ is strictly convex if and only if
 $\lambda_h$ is not null. This asymptotic configuration may arise due to 
 the definition of $\lambda_h$. Worth, in this case,  the function is 
 strictly decreasing and the minimal value is obtained when $p$ is the infinity.
 $\lambda_h$ is not null. This asymptotic configuration may arise due to 
 the definition of $\lambda_h$. Worth, in this case,  the function is 
 strictly decreasing and the minimal value is obtained when $p$ is the infinity.

+Thus, the method  follows its iterative calculus
+with an arbitrarely large value for $P_{sh}^{(k)}$. This leads to 
+a convergence which is dramatically slow down.
+

 
 To prevent this configuration, we replace the objective function given 
 in equation~(\ref{eq:obj2}) by 
 
 To prevent this configuration, we replace the objective function given 
 in equation~(\ref{eq:obj2}) by 


@@ -22,7 +26,7 @@ in equation~(\ref{eq:obj2}) by
 + \delta_p\sum_{h \in V }P_{sh}^{\frac{8}{3}}.
 \label{eq:obj2p}
 \end{equation}
 + \delta_p\sum_{h \in V }P_{sh}^{\frac{8}{3}}.
 \label{eq:obj2p}
 \end{equation}

-In this equation we have first introduced new regularisation factors
+In this equation we have first introduced new regularization factors

 (namely $\delta_x$, $\delta_r$, and $\delta_p$)
 instead of the sole $\delta$.  
 This allows to  further separately study the influence of each factor.
 (namely $\delta_x$, $\delta_r$, and $\delta_p$)
 instead of the sole $\delta$.  
 This allows to  further separately study the influence of each factor.


@@ -46,4 +50,26 @@ Provided $p^{5/3}$ is replaced by $P$, we have a quadratic function
 which is strictly convex, for any value of $\lambda_h$ since the discriminant 
 is positive. 
 
 which is strictly convex, for any value of $\lambda_h$ since the discriminant 
 is positive. 
 

-  
\ No newline at end of file
+This proposed enhancement has been evaluated as follows:  
+10 thresholds $t$, such that $1E-5 \le t \le 1E-3$, have 
+been selected and for each of them,  
+10 random configurations have been generated.
+For each one, we store the 
+number of iterations which is sufficient to make the dual 
+function variation smaller than this given threshold with 
+the two approaches: either the original one ore the
+one which is convex guarantee.
+
+The Figure~\ref{Fig:convex} summarizes the average number of convergence 
+iterations for each treshold value. As we can see, even if this new 
+enhanced method introduces new calculus, 
+it speeds up  the algorithm and guarantees the convexity, 
+and thus the convergence.
+\begin{figure*}
+\begin{center}
+\includegraphics[scale=0.5]{convex.png}
+\end{center}
+\caption{Original Vs Convex Guarantee Approaches}\label{Fig:convex}
+\end{figure*} 
+
+