Changeset 2774

Show
Ignore:
Timestamp:
10/30/09 13:32:47 (4 weeks ago)
Author:
jaein
Message:
 
Location:
HydroWatch/Tim/doc/ipsn10
Files:
2 modified

Legend:

Unmodified
Added
Removed

  • HydroWatch/Tim/doc/ipsn10/sec_eval.tex

    r2768 r2774  
    8383 
    8484 
    85 \subsubsection{Configuration} 
    86  
    87 For comparative analysis, we consider two related works: 
    88 Vigorito \cite{vigorito07} and Hsu \cite{hsu06} . Before we run each 
    89 algorithm, we configured each algorithm so that it produced 
    90 the best possible results.   
    91 Figure~\ref{fig:vigorito_algorithm} shows 
    92 that the performance of Vigorito's algorithm varies a lot  
    93 depending on the input parameters $\alpha$ and $\beta$: 
    94 \begin{itemize} 
    95 \item $\alpha$ : exponential moving average factor 
    96 for the historical average of duty-cycle ($\overline{\mu}$). 
    97 $\alpha$ is set to 0.01, where $|\overline{E_{tra}}|$  
    98 is minimized to 0.510 and zerodays is kept to 0.   
    99 \item $\beta$ : exponential moving average factor 
    100 between the duty-cycle calculation ($\mu$) and 
    101 its historical average ($\overline{\mu}$). 
    102 $\beta$ is set to 0.5, where $|\overline{E_{tra}}|$  
    103 is minimized  to 0.028. 
    104 \end{itemize} 
    105 %Thus, we set the input parameters as follows:  
    106 %$\beta$ = 0.5 and $\alpha$ = 0.01.  
     85%\subsubsection{Configuration} 
     86% 
     87%For comparative analysis, we consider two related works: 
     88%Vigorito \cite{vigorito07} and Hsu \cite{hsu06} . Before we run each 
     89%algorithm, we configured each algorithm so that it produced 
     90%the best possible results.   
     91%Figure~\ref{fig:vigorito_algorithm} shows 
     92%that the performance of Vigorito's algorithm varies a lot  
     93%depending on the input parameters $\alpha$ and $\beta$: 
    10794%\begin{itemize} 
    108 %\item $\beta$ : At $\beta = 0.5$, $|\overline{E_{tra}}|$  
     95%\item $\alpha$ : exponential moving average factor 
     96%for the historical average of duty-cycle ($\overline{\mu}$). 
     97%$\alpha$ is set to 0.01, where $|\overline{E_{tra}}|$  
     98%is minimized to 0.510 and zerodays is kept to 0.   
     99%\item $\beta$ : exponential moving average factor 
     100%between the duty-cycle calculation ($\mu$) and 
     101%its historical average ($\overline{\mu}$). 
     102%$\beta$ is set to 0.5, where $|\overline{E_{tra}}|$  
    109103%is minimized  to 0.028. 
    110 %\item $\alpha$ : At $\alpha = 0.01$, $|\overline{E_{tra}}|$  
    111 %is minimized to 0.510 while keeping the zerodays to 0.   
    112 %\end{itemize}  
    113 In a similar way, we set the moving average factor $\alpha$  
    114 that is used to predict the daily harvested solar energy of  
    115 the Hsu's algorithm. We set the $\alpha$ so that the  
    116 algorithm minimizes the difference from the target energy  
    117 level, which is at $\alpha = 0.75$.   
    118 %Figure~\ref{fig:hsu_algorithm} shows that the 
    119 %performance of Hsu's algorithm depends on the exponential  
    120 %moving average factor $\alpha$ that is used to predict  
    121 %the daily harvested solar energy $\overline{E_s}$ 
    122 %from the historical measurement $E_s$. We set the $\alpha$  
    123 %so that the algorithm minimizes the difference from the  
    124 %target energy level, which is at $\alpha = 0.75$.  
    125  
    126 \begin{figure} 
    127 %\begin{tabular}{|p{0.45\textwidth}|} 
    128 %\hline 
    129 %\begin{scriptsize} 
    130 %\begin{algorithmic}[1] 
    131 %%\STATE $\theta$ $\leftarrow$ $[2,-1,1]$ 
    132 %%\STATE $B$ $\leftarrow$ $B_{init} / B_{max}$ 
    133 %%\STATE $B^*$ $\leftarrow$ $B_{target} / B_{max}$ 
    134 %%\STATE $\rho$, $\mu$, $\overline{\mu}$ $\leftarrow$ $DC_{init}$ 
    135 %%%\STATE $\mu$ $\leftarrow$ $DC_{init}$ 
    136 %%%\STATE $\overline{\mu}$ $\leftarrow$ $DC_{init}$ 
    137 %%\STATE $\phi$ $\leftarrow$ $[B, \mu, -B^*]$ 
    138 %\LOOP 
    139 %  \STATE $B$ $\leftarrow$ $Bat / B_{max}$ 
    140 %  \STATE $\theta$ $\leftarrow$ $\theta + \frac{\mu}{\phi \cdot \phi} \times (B - \phi \cdot \theta) \times \phi$  
    141 %  \STATE $\mu$ $\leftarrow$ $\min(1, \max(0, \frac{B^* - \theta(1) \times B + \theta(3) \times B^*}{\theta(2)}))$ 
    142 %  \STATE $\phi$ $\leftarrow$ $[B, \mu, -B^*]$ 
    143 %  \STATE $\overline{\mu}$ $\leftarrow$ $\overline{\mu} + \alpha \times (\mu - \overline{\mu})$ 
    144 %  \STATE $\rho$ $\leftarrow$ $\beta \times \mu + (1 - \beta) \times \overline{\mu}$ 
    145 %  \STATE $1/T_{samp}$ $\leftarrow$ $(1-\rho)/T_{sampmax} + \rho/T_{sampmin}$ 
    146 %  \STATE $1/T_{report}$ $\leftarrow$ $(1-\rho)/T_{reportmax} + \rho/T_{reportmin}$ 
    147 %\ENDLOOP  
    148 %\end{algorithmic} 
    149 %\end{scriptsize} 
    150 %\\ 
    151 %\hline 
    152 %\end{tabular} 
    153   \begin{tabular}{cc} 
    154         \includegraphics[width=0.23\textwidth]{fig/vigoritto_case3} &  
    155         \includegraphics[width=0.23\textwidth]{fig/vigoritto_case2} \\ 
    156     \end{tabular} 
    157 \caption{Configuring the input parameters $\alpha$, $\beta$ 
    158 for Vigorito's algorithm}  
    159 \label{fig:vigorito_algorithm} 
    160 \end{figure} 
    161  
    162 %\begin{figure*}[ht] 
    163 %    \centering 
    164 %  \begin{tabular}{ccc} 
    165 %        \includegraphics[width=0.25\textwidth]{fig/vigoritto_case3} &  
    166 %        \includegraphics[width=0.25\textwidth]{fig/vigoritto_case2} & 
    167 %        \includegraphics[width=0.25\textwidth]{fig/hsu_case1} \\  
    168 %        (a) $\alpha$ of Vigorito & 
    169 %        (b) $\beta$ of Vigorito &  
    170 %        (c) $\alpha$ of Hsu \\  
     104%\end{itemize} 
     105%%Thus, we set the input parameters as follows:  
     106%%$\beta$ = 0.5 and $\alpha$ = 0.01.  
     107%%\begin{itemize} 
     108%%\item $\beta$ : At $\beta = 0.5$, $|\overline{E_{tra}}|$  
     109%%is minimized  to 0.028. 
     110%%\item $\alpha$ : At $\alpha = 0.01$, $|\overline{E_{tra}}|$  
     111%%is minimized to 0.510 while keeping the zerodays to 0.   
     112%%\end{itemize}  
     113%In a similar way, we set the moving average factor $\alpha$  
     114%that is used to predict the daily harvested solar energy of  
     115%the Hsu's algorithm. We set the $\alpha$ so that the  
     116%algorithm minimizes the difference from the target energy  
     117%level, which is at $\alpha = 0.75$.   
     118%%Figure~\ref{fig:hsu_algorithm} shows that the 
     119%%performance of Hsu's algorithm depends on the exponential  
     120%%moving average factor $\alpha$ that is used to predict  
     121%%the daily harvested solar energy $\overline{E_s}$ 
     122%%from the historical measurement $E_s$. We set the $\alpha$  
     123%%so that the algorithm minimizes the difference from the  
     124%%target energy level, which is at $\alpha = 0.75$.  
     125% 
     126%\begin{figure} 
     127%%\begin{tabular}{|p{0.45\textwidth}|} 
     128%%\hline 
     129%%\begin{scriptsize} 
     130%%\begin{algorithmic}[1] 
     131%%%\STATE $\theta$ $\leftarrow$ $[2,-1,1]$ 
     132%%%\STATE $B$ $\leftarrow$ $B_{init} / B_{max}$ 
     133%%%\STATE $B^*$ $\leftarrow$ $B_{target} / B_{max}$ 
     134%%%\STATE $\rho$, $\mu$, $\overline{\mu}$ $\leftarrow$ $DC_{init}$ 
     135%%%%\STATE $\mu$ $\leftarrow$ $DC_{init}$ 
     136%%%%\STATE $\overline{\mu}$ $\leftarrow$ $DC_{init}$ 
     137%%%\STATE $\phi$ $\leftarrow$ $[B, \mu, -B^*]$ 
     138%%\LOOP 
     139%%  \STATE $B$ $\leftarrow$ $Bat / B_{max}$ 
     140%%  \STATE $\theta$ $\leftarrow$ $\theta + \frac{\mu}{\phi \cdot \phi} \times (B - \phi \cdot \theta) \times \phi$  
     141%%  \STATE $\mu$ $\leftarrow$ $\min(1, \max(0, \frac{B^* - \theta(1) \times B + \theta(3) \times B^*}{\theta(2)}))$ 
     142%%  \STATE $\phi$ $\leftarrow$ $[B, \mu, -B^*]$ 
     143%%  \STATE $\overline{\mu}$ $\leftarrow$ $\overline{\mu} + \alpha \times (\mu - \overline{\mu})$ 
     144%%  \STATE $\rho$ $\leftarrow$ $\beta \times \mu + (1 - \beta) \times \overline{\mu}$ 
     145%%  \STATE $1/T_{samp}$ $\leftarrow$ $(1-\rho)/T_{sampmax} + \rho/T_{sampmin}$ 
     146%%  \STATE $1/T_{report}$ $\leftarrow$ $(1-\rho)/T_{reportmax} + \rho/T_{reportmin}$ 
     147%%\ENDLOOP  
     148%%\end{algorithmic} 
     149%%\end{scriptsize} 
     150%%\\ 
     151%%\hline 
     152%%\end{tabular} 
     153%  \begin{tabular}{cc} 
     154%        \includegraphics[width=0.23\textwidth]{fig/vigoritto_case3} &  
     155%        \includegraphics[width=0.23\textwidth]{fig/vigoritto_case2} \\ 
    171156%    \end{tabular} 
    172 %    \caption{Configuring input parameters} 
    173 %    \label{fig:vigorito_parameters} 
    174 %\end{figure*} 
    175  
    176 %\begin{figure} 
    177 %\centering 
    178 %\includegraphics[width=0.25\textwidth]{fig/hsu_case1}  
    179 %\caption{Configuring the input parameter $\alpha$ for 
    180 %Hsu's algorithm} 
    181 %\label{fig:hsu_algorithm} 
    182 %\end{figure} 
    183  
    184  
    185  
    186 \subsubsection{Simulation Results} 
    187  
    188 To compare the performance of Vigorito's algorithm 
    189 with ours, we consider the following metrics: 
     157%\caption{Configuring the input parameters $\alpha$, $\beta$ 
     158%for Vigorito's algorithm}  
     159%\label{fig:vigorito_algorithm} 
     160%\end{figure} 
     161% 
     162%%\begin{figure*}[ht] 
     163%%    \centering 
     164%%  \begin{tabular}{ccc} 
     165%%        \includegraphics[width=0.25\textwidth]{fig/vigoritto_case3} &  
     166%%        \includegraphics[width=0.25\textwidth]{fig/vigoritto_case2} & 
     167%%        \includegraphics[width=0.25\textwidth]{fig/hsu_case1} \\  
     168%%        (a) $\alpha$ of Vigorito & 
     169%%        (b) $\beta$ of Vigorito &  
     170%%        (c) $\alpha$ of Hsu \\  
     171%%    \end{tabular} 
     172%%    \caption{Configuring input parameters} 
     173%%    \label{fig:vigorito_parameters} 
     174%%\end{figure*} 
     175% 
     176%%\begin{figure} 
     177%%\centering 
     178%%\includegraphics[width=0.25\textwidth]{fig/hsu_case1}  
     179%%\caption{Configuring the input parameter $\alpha$ for 
     180%%Hsu's algorithm} 
     181%%\label{fig:hsu_algorithm} 
     182%%\end{figure} 
     183% 
     184% 
     185% 
     186%\subsubsection{Simulation Results} 
     187 
     188For comparative analysis, we consider two related works 
     189on energy-harvesting-aware duty-cycling algorithms: 
     190Vigorito \cite{vigorito07} and Hsu \cite{hsu06} .  
     191To compare the performance of these algorithms with ours, 
     192we consider the following metrics:  
    190193\begin{itemize} 
    191194\item Utility : defined in formula (2). 

  • HydroWatch/Tim/doc/ipsn10/sec_related.tex

    r2717 r2774  
    4444global synchronization protocols \cite{ye02infocom,tmac03sensys}  
    4545and keep the duty-cycle very small. Dozer uses a tree-routing. 
    46 As well as routing data packtes over multiple hops, this allows 
    47 the protocol pass the per-hop TDMA schedule without building 
     46This allows the protocol pass the per-hop TDMA schedule without building 
    4847additional path for the control traffic.   
    4948Koala \cite{koala08ipsn} provides a lower-duty cycling link-level 
    … …  
    5554the receiver by sending an acknowledgement and sends a data packet.  
    5655This mechanism generates much less traffic than the LPL which transmits  
    57 a sequence of packetized preambles on the sender side. Thus, LPP can be  
    58 adopted in a broadcast environment, where multiple nodes can send and  
    59 receive at the same time.  
     56a sequence of packetized preambles on the sender side.  
     57%Thus, LPP can be  
     58%adopted in a broadcast environment, where multiple nodes can send and  
     59%receive at the same time.  
    6060 
    6161\subsection{Solar Energy Harvesting and Adaptive Energy Management} 
    … …  
    9292Spatial adaptive sampling adjusts sampling density depending 
    9393on how abruptly a physical phenomenon changes, and it can be  
    94 further divided into \textit{mobile adaptive sampling} 
     94divided into \textit{mobile adaptive sampling} 
    9595\cite{batalin04sensys,zhang04iros,singh06ipsn}, which adjusts  
    9696sampling density by physically moving a sensor over a range of space,  
    … …  
    9898which selectively chooses representative sensor nodes among the  
    9999sensor nodes deployed. Temporal adaptive sampling adjusts sampling  
    100 rate over time, and it can be further divided into \textit{adaptive  
     100rate over time, and it can be divided into \textit{adaptive  
    101101sensor sampling} \cite{batalin04sensys} or \textit{adaptive sensor  
    102102filtering} \cite{jain04dmsn,santini06inss} depending on where the  
    … …  
    111111it optimizes utility (high-resolution data report rate) with 
    112112the varying energy availability and the given constraints, 
    113 but it has several differences. 
    114 First, our work considers varying energy availability of a solar 
    115 energy harvesting system, which requires estimation on energy  
    116 variability due to daily, seasonal 
    117 and meteorological variance. Whereas, Lance assumes 
    118 a simple model of non-rechargeable battery and linear 
    119 energy consumption rate. 
    120 Second, both our work and Lance provides a type of quality of service. 
    121 With the contraints of energy availability and lifetime target, 
    122 Lance prefers data values that is higher than preset threshold. 
    123 With the case of Lance, this is a good strategy because it prefers 
    124 more meaningful reading while dropping less meaningful data. 
    125 However, this prioritization does not hold in a general case. 
    126 Our work provides quality of service by providing two-levels 
    127 of frequency: low-resolution summary frequency and 
    128 high-resolution data report frequency. Our work guarantees that 
    129 summary of data is reported at a preset interval while 
    130 varies the high-resolution data report frequency based 
    131 on energy availability. 
     113but it is different in energy model and data filtering method.   
     114As for the energy model, our work considers a solar energy 
     115harvesting system, which requires estimation on energy  
     116variability due to daily, seasonal and meteorological variance.  
     117Whereas, Lance assumes a simple model of non-rechargeable battery  
     118and linear energy consumption rate. 
     119As for the data filtering method, Lance selects data values 
     120higher than preset threshold rather than evenly selecting the values.   
     121This method can be effective with the case of Lance but is not  
     122applicable in general. 
     123 
     124 
     125%Second, both our work and Lance provides a type of quality of service. 
     126%With the contraints of energy availability and lifetime target, 
     127%Lance prefers data values that is higher than preset threshold. 
     128%With the case of Lance, this is a good strategy because it prefers 
     129%more meaningful reading while dropping less meaningful data. 
     130%However, this prioritization does not hold in a general case. 
     131%Our work provides quality of service by providing two-levels 
     132%of frequency: low-resolution summary frequency and 
     133%high-resolution data report frequency. Our work guarantees that 
     134%summary of data is reported at a preset interval while 
     135%varies the high-resolution data report frequency based 
     136%on energy availability. 
    132137 
    133138