Sensor models and localization algorithms for sensor networks based on received signal strength
 Fredrik Gustafsson^{1}Email author,
 Fredrik Gunnarsson^{1} and
 David Lindgren^{2}
https://doi.org/10.1186/16871499201216
© Gustafsson et al; licensee Springer. 2012
Received: 31 December 2010
Accepted: 16 January 2012
Published: 16 January 2012
Abstract
Received signal strength (RSS) can be used in sensor networks as a ranging measurement for positioning and localization applications. This contribution studies the realistic situation where neither the emitted power nor the power law decay exponent be assumed to be known. The application in mind is a rapidly deployed network consisting of a number of sensor nodes with lowbandwidth communication, each node measuring RSS of signals traveled through air (microphones) and ground (geophones). The first contribution concerns validation of a model in logarithmic scale, that is, linear in the unknown nuisance parameters (emitted power and power loss constant). The parameter variation is studied over time and space. The second contribution is a localization algorithm based on this model, where the separable least squares principle is applied to the nonlinear least squares (NLS) cost function, after which a cost function of only the unknown position is obtained. Results from field trials are presented to illustrate the method, together with fundamental performance bounds. The ambition is to pave the way for sensor configuration design and more thorough performance evaluations as well as filtering and target tracking aspects.
Keywords
1 Introduction
Target localization based on the target's emitted energy is an attractive option in large, wireless sensor networks:

Simple and passive (no energy output) sensors like microphones and geophones can be used.

Requirements on network synchronization are moderate.

Data fusion requires limited communication bandwidth.
By sampling the energy as a measure of the Received Signal Strength (RSS) at geographically distributed locations and by modeling the energy decay as a function of targetsensor distance, the location of the target can be inferred. This paper focuses on centralized acoustic and seismic source localization, which is interesting to use as a part of surveillance systems for power plant protection, airport security, border control, and similar. However, the models and algorithms are applicable to general target localization based on emitted energy from the target.
Energy source localization is in focus here, but the reverse problem of navigation of one sensor ("sink") from several beacons ("sources") with known position is also covered by reversing the role of emitters and sensors. Therefore, the object to be located will be referred to as the target in this paper. An underlying assumption is that communication constraints between the sensor units make any algorithm based on the signal waveform (like coherent detection) infeasible. Communication only allows for sending RSS measurements to other sensor units.
Localization from received signal energy is of course a fairly wellstudied problem, see the surveys [1–3] and the papers [4, 5], though the major part of literature addresses the related problem of localization from time of arrival (TOA) and timedifference of arrival (TDOA) measurements. Also, the standard localization application concerns radio networks, but localization in acoustic networks bears much in common. While TOA measures range and TDOA range differences computed from propagation time, energybased localization utilizes the power decay of the involved signals.
Based on the distance power law model, the received root mean square (RMS) signal power expressed in decibels (dB) is assumed to be proportional to the logarithm of distance, and this is the main difference to timebased localization approaches. Dedicated approaches to this problem assume that the constant of proportionality (power law decay exponent) is known [4–6] or include the energy measurements as a general nonlinear relation [3]. Several ad hoc methods to eliminate nuisance parameters have been proposed in this context, including taking pairwise differences or ratios of observations.
A least squares solution for energybased methods can be found in [4], in which the power law model is verified. However, no investigations regarding proper noise models were conducted. Maximumlikelihood (ML) estimators are considered in [6, 7] based on the same power law model, but with a fixed and known power law decay exponent. The same holds for [8], but the focus is on least squares based approaches. These works consider a centralized situation, where all measurements are processed at the same location. Distributed ML is addressed in [9], where the authors consider both the power law decay exponent and source energy as unknowns.
In [7], an approach to localization based on a model in linear energy scale is presented, where the power law decay exponent is fixed to 2. In [10, 11], a similar model was used, but in logarithmic energy scale. The model is referred to as the log range linear model, where all environmental parameters including power law decay exponent appear linearly. This is of course a great advantage in estimation. The first purpose of this contribution is to use measurements from extensive field tests to validate the log range linear model.
The second contribution is to extend the theory of RSS based localization using an approach where the power law decay exponent and emitted power are explicitly removed from a set of RSS measurements using the separable least squares principle, after which the resulting problem is nonlinear in target state parameters only. This leads to a standard lowdimensional nonlinear least squares (NLS) problem, where efficient numerical algorithms exist. Algorithms of different complexity and performance are outlined for this framework. Tracking algorithms are also described, which are based on stating the localization NLS problem formulation as the measurement relation in an extended Kalman filter.
The fundamental performance bound implied by the CramérRao lower bound enables efficient analysis of sensor network architecture, management, and resource allocation. This bound has been analyzed thoroughly in the sensor network literature, primarily for TOA, TDOA, and angleofarrival (AOA), [1–3], but also for RSS [12, 13] and with specific attention to the impact from nonlineofsight [14, 15]. The nonlineofsight signal propagation is also related to multipath signal propagation, where the signal is reflected and is received as multiple copies, essentially as multiple nonlineofsight signals replicas. Numerical explicit algorithms and CramérRao lower bounds (CRLB) for both stationary and moving target are derived for the NLS problem formulation.
In Section 2, the RSS model is introduced and compared first to some previously proposed models and then to some simplified models useful for detection purposes. Section 3 validates the model and its assumptions using extensive field tests with acoustic and seismic signals. Section 4 presents a nonlinear least squares (NLS) framework for localization, where the separable least squares (SLS) principle can be used to eliminate nuisance parameters. Localization and tracking algorithms based on this framework are overviewed in Section 5, illustrated with selected results from field trials. Finally, Section 7 concludes the paper.
2 RSS measurement model
It is assumed that the received signal strength (RSS) measured at each sensor is proportional to the targetsensor distance to the power of a constant or a parameter, the distance power law. For a motivating example, the signal from a microphone (measured in Volts) is ideally proportional to the sound pressure (in Pascal), which in turn decays inversely proportional to the sound source distance. This holds for punctual sources and freespace propagation. By identifying the RSS with the mean square of the received signal, the decay (for punctual sources in free space) is thus expected to decay inversely proportional to the distance square. However, in the nonideal case, factors like reflection, diffraction, and refraction influence the RSS decay in a way that generally is difficult or expensive to predict. The approach here is to keep the distance power law for its simplicity, but allow for the decay exponent to adapt to the current situation. Thus, the exponent is considered as an unknown parameter.
We reason analogously in the seismic case, although the seismic wave propagation is even more involved to predict than is the acoustic. The particle amplitude of seismic surface waves (Rayleigh and Love waves) ideally decays inversely proportional to the square root distance, which is in agreement with the distance power law, although distinct from the acoustic case.
The acoustic and seismic distance power laws with unknown exponents will here be validated on data collected in fairly open terrain. Generally, the distance power law is probably too simplistic to work well in dense urban environments, where the RSS results from a superposition of multiple wave reflections with different path distances. The urban nonlineofsight case will, however, not be treated in this text.
The RSS value itself is here computed by a first optional prefiltering step of the raw sensor signal, then an averaging of the squared magnitude over a time window to obtain a downsampled signal. The signals are typically sampled with 14 kHz, and the final RSS value is obtained with one or a few Hz sampling frequency. The prefilter only passes signals in frequency bands of interest in the application, and the averaging reduces the variations in the RSS estimates.
2.1 Notation
The following notation will be used throughout the paper:

The sensor network consists of M sensor nodes. These are located at p_{ m }, where m is used for sensor index.

Each sensor node is equipped with several different sensors, and i indicates the sensor types. In the field tests, i = 1 corresponds to a microphone and i = 2 to a geophone.

There is one moving target with position x_{ n }at time n. There are N time instants in each experiment.

y_{ i,n,m }is the received energy, RSS, at sensor node m, observed by sensor type i, and averaged over time window n.

e_{ i,n,m }is additive noise with variance ${\sigma}_{n,m}^{2}$.

θ_{1,i,n}denotes the reference received energy for sensor type i if it would have been placed 1 m from the target, averaged over time window n.

θ_{2,i,m}denotes the attenuation or measurement error bias of sensor node m at sensor type i.

θ_{3,i,n,m}denotes the power law decay exponent, which may vary with sensor type, time, and space.

θ(n, m) = [θ_{1,n}, θ_{2,m}, θ_{3,n,m}]^{ T }gathers the parameters in a parameter vector
A convention is that energy variables are primarily defined in logarithmic energy scale, while a bar on a variable indicates values in linear scale. Hence, $y=\text{log}\left(\u0233\right)$.
2.2 Parametric model
Note that (3) and (1) are identical in the noisefree case when θ_{3,n,m}= 2 and θ_{2,m}= 0 (no sensor biases).
It is rather obvious that the models above have a limited scope in the range ∥p_{ m } x_{ n }∥. First, when the target distance tends to zero, the models predict infinite RSS (in log scale). Beside saturation in the sensors, there are nearfield and other effects that limit the validity for close distances. Second, when the range tends to infinity, the models (2) and (3) both predict negative infinite RSS. For large distances, the background noise will dominate the target signal in these models.
2.3 Simplified models
The problem is primarily to find the target locations x_{ n }, n = 1, 2, ..., N. The parameters in θ(n, m) are considered nuisance. Nevertheless, the dimension of θ (n and m omitted for notational simplicity) is 2(M + N + NM) and the number of target coordinate figures, assuming a moving target, is 3N. However, there are only 2NM observations, so without any further assumptions, the system of equations is underdetermined, and the localization problem not well defined. The reason for initially describing this overgeneral model is that it will not only be used for localization, but also for model validation, where a ground truth could be used in place of otherwise unknown figures. Thus, the following submodels are defined, which all correspond restricted versions of (3):
M0 θ = 0, that is, there is no targetdependent relation at all.
M1 θ_{3,i,n,m}= 0 for all i, n, m. This bias model just compensates for sensor bias and common target energy for all sensors. That is, there is no range dependence of the target.
M2 θ_{3,i,n,m}= θ_{3,i}for all n,m,i. This log range model assumes that the path loss is a global timeinvariant constant.
Of these models, only M2 (with 5N+2M+2 unknowns) can be used for actual localization. M0 and M1 are analyzed to provide a reference to which the M2 performance and accuracy will be compared. Thus, the objective in Section 3 is to validate M2 by showing that it gives significantly better predictions of the measured y_{ i,n,m }than M0 and M1. Moreover, even though the models M0 and M1 are not suitable for localization, they can still be useful for target detection, representing the hypothesis that no target is present. The more detailed interpretation is that M0 corresponds to no observed signal and no nodespecific bias, and M1 no observed signal, but a nodespecific bias.
3 Model validation
Before analyzing localization algorithms, the energy decay model will be validated on real data. Our data set contains GPS position of the target, so the relation between RSS and targetsensor distance can be analyzed given known distances.
3.1 LS estimation of θ
The principle for model validation is to use a known network configuration $\mathcal{P}$ and a known trajectory $\mathcal{X}$ to estimate the parameters, and as performance indicators compare (i) the model residuals $y{\Phi}^{T}\left(\mathcal{X},\mathcal{P}\right)\widehat{\theta}$, (ii) the parameters with their respective confidence intervals, and the (iii) obtained RMSE for each model.
3.2 Single sensor experiments
The positions of the vehicle and sensor are known perfectly in this experiment.
The results and conclusions from these figures, which are also supported by data from the sensor network field trials in Section 3.3, are as follows:

Figure 2 indicates that the microphone is subject to more noise or rather variations than the geophone, provided that the proposed models are relevant. The variations are probably due to wind gusts as well as fading effects when the ground reflected wave interferes with the lineofsight wave. Such fading effects have been analyzed thoroughly for radio channel using the tworay model, see for example [18].

Figure 2 indicates that the RSS is slightly skewed and more energy is received when the vehicle is moving away from the sensor compared to when it is moving toward the sensor at the same distances from the sensor. This is explained by the fact that there is more sound coming from the back of the vehicle, the exhaust pipe end, compared to the front.

Figure 3 shows that the RSS is linearly dependent of the log range, which verifies model M2. There is a slight nearfield effect for the microphone, in that the RSS value saturates for short distances.

From Figure 3, the noise contribution as well as the fading effects appears fairly independent of range in logarithmic scale, which would confirm the assumption that noise is additive to the logarithmic RSS measurements.
3.3 Sensor network experiments

A motorcycle (MC).

A fourwheeled allterrain vehicle (FW).
The two data sets denoted MC and FW contain different number of samples N. Reference positions of the targets are measured with differential GPS (DGPS) with sampling rate 1 Hz.
For least squares estimation, sensor data from the target positions in the time interval (20, 30) seconds are used. This is due to the rapidly decreasing signal quality as a function of range and that the purpose here is sensor model validation rather than tracking. The results and conclusions are as follows:

Table 1 shows the RMSE value of the received logarithmic energy for model M0 (raw data), model M1 and model M2, respectively. The RMSE is significantly smaller with log range in the model.
RMSE (dB) for no model (M0), bias model (M1), and log range model (M2), respectively
RMSE0  RMSE1  RMSE2  

FWmic  32.5  3.9  3.2 
MCmic  30.1  4.7  4.1 
FWgeophone  34.1  3.3  2.2 
MCgeophone  35.7  4.1  2.8 

Table 2 shows the estimated log range parameter ${\widehat{\theta}}_{3}$ together with its standard deviation (square root of last diagonal element of P_{ θ }). The standard deviation is orders of magnitude smaller than the parameter estimate, indicating an accurate estimate. The conclusion is that no generic constant, as for instance 2, should be used.
Estimated power law exponent θ_{3,i}with its standard deviation using the log range model (M2)
θ _{3}  std  

FWmic  2.6  0.02 
MCmic  2.6  0.04 
FWgeophone  2.8  0.01 
MCgeophone  3.4  0.02 

Figure 5 shows the estimated sensor bias. For instance, microphone number 2 has a significant bias in both data sets. The sensors themselves are factory calibrated, but the deployment may give cause to a systematic difference. These terms also capture individual background noise and lineofsight/nonlineofsight issues. That is, the sensor bias term is needed. On the other hand, the timevarying offset is not significantly different from zero and can be neglected for a single target passage.

Figure 6 shows the spectrum for the sensor signal between 23 and 24 s. Each vehicle seems to have a characteristic signature, with one fundamental frequency, and a lot of harmonics.

The model residuals from both microphones and geophones are illustrated as smoothed histograms in Figure 7 and compared to Gaussian approximations. The Gaussian noise assumption is apparently quite realistic. The noise standard deviation can be set to σ ∈ (2, 3).
4 Eliminating nuisance θ
The goal in this section is to eliminate the nuisance parameters θ, including the power law exponent, the emitted energy, and optionally the unknown noise variances σ.
In this section, we assume that the noise variance σ^{2} is known. Also note that the parameter θ_{1,i,n}, representing timevarying background noise in (4e) is irrelevant in this snapshot formulation, since it cannot be distinguished from the sensorvarying background noise θ_{2,i,m}.
4.1 Separable Least Squares
This is larger than the measurement error variance σ^{2} alone.
Note that the real observation y_{ m }is seen as a known input, and the virtual observation ${\u0177}_{m}$ is always zero. This signal model depends on x only, and the new measurement error is additive white Gaussian with unity variance.
4.2 SNLS formulation
The cost function V_{SNLS} in (9) is similar to the cost function V in (6), except that the model prediction error variance λ_{ m }is considered in the former, while the sensor measurement error variance σ^{2} is considered in the latter. Hence, the new weighting in the sum of least squares accounts for both measurement noise and the estimation uncertainty in the nuisance parameters. Typically, far away sensor nodes m get larger uncertainty in the parameters and thus automatically a smaller weight in the criterion.
4.3 Sensor noise variance estimation
where the normalization with N  2 accounts for the degrees of freedom lost by the minimization and is needed to get an unbiased variance estimate. The last equality is a consequence of the LS theory and will be used in the NLS formulation below.
5 Localization algorithms
We have here omitted the dependence of the original observation in (12a). Here, $\widehat{\theta}\left(x\right)$ is given in (7a), c_{ m }(x) in (7d), $h\left({c}_{m}\left(x\right),\widehat{\theta}\right)$ in (7c), and λ(x) in (7d) (using the measurement variance estimate $\hat{{\sigma}^{2}}\left(x\right)$ from (10)). The purpose in this section is to outline possible implementation strategies.
5.1 Estimation criteria
Optimization criteria V(x) for estimating position x from uncertain measurements y = h(x) + e
NLS  V^{NLS}(x) = (y  h(x))^{ T }R^{1}(x)(y  h(x)) 

GML  V^{GML}(x) = (y  h(x))^{ T }R^{1}(x)(y  h(x))+log detR(x) 
ML  V^{ML}(x) = log p_{e}(y  h(x)) 
5.2 Eliminating the Noise Variance
The logarithm intuitively decreases the difference in weighting between the different sensor types compared to the case of known noise variances in (9b).
5.3 Optimization
Estimation algorithms applicable to optimization criteria in Table 3
Gridbased  ${\widehat{x}}^{o}=\text{arg}\underset{x\in \left\{{x}^{\left(1\right)},{x}^{\left(2\right)},\dots ,{x}^{\left(L\right)}\right\}}{\text{min}}V\left(x\right)$ 

Steepest descent  ${\widehat{x}}_{j}^{o}={\widehat{x}}_{j1}^{o}{\mu}_{k}{\nabla}_{x}\left[V\left(x\right)\right]$ 
NewtonRaphson  ${\widehat{x}}_{j}^{o}={\widehat{x}}_{j1}^{o}+{\mu}_{k}{\left({\nabla}_{x}\left[{\nabla}_{x}^{T}\left[V\left(x\right)\right]\right]\right)}^{1}{\nabla}_{x}\left[V\left(x\right)\right]$ 
5.3.1 GridBased optimization
where ${\phi}_{\ell}={\left[{\phi}_{\ell}^{1},{\phi}_{\ell}^{2},{\phi}_{\ell}^{3}\right]}^{T}$ and η = [η^{1}, η^{2}, η^{3}]^{ T }. Note that some rows of vec(P_{ x }) will be identical since P_{ x }is symmetric and duplicates shall therefore be removed from η, as well as corresponding rows of φ_{ℓ}. This is thus a linear regression resulting in an overdetermined system if the number of local grid points are sufficiently large. Least squares provides the solution $\widehat{\eta}$ to (17), from which P_{ x }and x^{ o }can be derived.
5.3.2 Gradient derivation
The gradient H(x) = ∇_{ x }h(v) of the model with respect to the position is instrumental in several loss function minimization algorithms, and it is the purpose here to derive the necessary equations.
The gradient of the NLS loss function V(x)) becomes a function of the gradients of $h\left(x,\widehat{\theta}\right)$ and λ(x). These are all tedious but straightforward applications of the chain rule, not reproduced here. However, the gradient can be expressed as a closed expression based on the target location x and sensor locations p_{ m }.
6 Fundamental bounds
The Fisher Information Matrix (FIM) provides a fundamental estimation limit for unbiased estimators referred to as the CramérRao lower bound (CRLB) [23]. This bound has been analyzed thoroughly in the literature, primarily for AOA, TOA, and TDOA, [1–3], but also for RSS [12, 13] and with specific attention to the impact from nonlineofsight [14, 15].
where p is the twodimensional position vector and p_{e}(y  h(x)) the likelihood given the error distribution. For the SNLS model 8, J(x) is 2 × 2. Again, the gradient derivations are tedious but symbolic exercises not reproduced here.
The former information measure is additive as FIM itself, while the latter is an underestimation of the information useful when reasoning about whether the available information is sufficient or not. Note that in the Gaussian case with a diagonal measurement error covariance matrix, the trace of FIM is the squared gradient magnitude.
where x^{ o }denotes the true position. (22) holds for any unbiased estimate of $\widehat{x}$, although the right hand side is not necessarily attainable. Asymptotically in the number of sensor nodes, the ML estimate is $\widehat{x}~N\left({x}^{o},{J}^{1}\left({x}^{o}\right)\right)$ [24] and thus reaches this bound, but this may not hold for finite amount of data.
The right hand side of (22) gives, however, an idea of how suitable a given sensor configuration is for positioning. It can also be used for sensor network design. It should always be kept in mind though that this lower bound is quite conservative and relies on many assumptions.
If RMSE requirements are specified, it is possible to include more and more measurements in the design until (23) indicates that the amount of information is enough.
7 Conclusions
Conventional received signal strength (RSS)based algorithms as found in the literature of wireless or acoustic networks assume either that the emitted power is known or that the distance power law exponent is known from calibration. We have considered a network of microphone sensors that is rapidly deployed in an unknown environment where the distance power law exponent is unknown or may vary with time. Also, the emitted power is inherently unknown in the localization and tracking applications under consideration. For localization, both the emitted acoustic energy and the power law exponent are nuisance parameters unique for each target and sensor type, but constant over the sensor nodes.
The nonlinear least squares (NLS) algorithm offers a suitable framework for positioning in this kind of sensor networks, where the RSS measurements suffer from unknown emitted power and where also the environmental path loss constant is unknown. Marginalization of the nuisance parameters using the separable least squares principle leads to a NLS cost function of only two unknowns (horizontal position), where global gridbased methods can be used for minimization. Results from field trials confirm the usability of the proposed method. Hopefully, the provided framework can form a basis for subsequent target tracking and thorough performance evaluations.
Declarations
Acknowledgements
This work is funded by the VINNOVA supported Centre for Advanced Sensors, Multisensors and Sensor Networks, FOCUS, at the Swedish Defence Research Agency, FOI. The authors also want to thank two anonymous reviewers for very detailed comments and suggestions on improvements.
Authors’ Affiliations
References
 Patwari N, Hero A III, Perkins M, Correal N, O'Dea R: Relative location estimation in wireless sensor networks. IEEE Trans Signal Process 2003, 51(8):21372148. 10.1109/TSP.2003.814469View ArticleGoogle Scholar
 Gezici S, Tian Z, Giannakis B, Kobayashi H, Molisch A: Localization via ultrawideband radios. IEEE Signal Process Mag 2005, 22(4):7084.View ArticleGoogle Scholar
 Gustafsson F, Gunnarsson F: Possibilities and fundamental limitations of positioning using wireless communication networks measurements. IEEE Signal Process Mag 2005, 22: 4153.View ArticleGoogle Scholar
 Li D, Hu Y: Energybased collaborative source localization using acoustic microsensor array. J Appl Signal Process 2003, 321337.Google Scholar
 Huang Y, Benesty J, Elko B: Passive acoustic source localization for video camera steering. IEEE Conference on Acoustics, Speech and Signal Processing 2000.Google Scholar
 Chen LWHPCM, Liu Z, Zhang Z: Energybased position estimation of microphones and speakers for ad hoc microphone arrays. Proceedings of of IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, New Paltz, NY 2007.Google Scholar
 Sheng X, Hu YH: Maximum likelihood multiplesource localization using acoustic energy measurements with wireless sensor networks. IEEE Trans Signal Process 2005, 53(1):4453.MathSciNetView ArticleGoogle Scholar
 Meesookho UMC, Narayanan S: On energybased acoustic source localization for sensor networks. IEEE Trans Signal Process 2008, 56(1):365377.MathSciNetView ArticleGoogle Scholar
 Shi Q, He C: A new incremental optimization algorithm for MLbased source localization in sensor networks. IEEE Signal Process Lett 2008, 15: 4548.View ArticleGoogle Scholar
 Gustafsson F, Gunnarsson F: Localization in sensor networks based on log range observations. In Fusion 2007. Quebec, Canada; 2007.Google Scholar
 Gustafsson F, Gunnarsson F: Localization based on observations linear in log range. In International Federation of Automatic Control (IFAC) World Congress. IFAC, Seoul; 2008.Google Scholar
 Koorapaty H: Barankin bound for position estimation using received signal strength measurements. Proceedings of IEEE Vehicular Technology Conference, Milan, Italy 2004.Google Scholar
 Qi Y, Kobayashi H: On relation among time delay and signal strength based geolocation methods. Proceedings of IEEE Global Telecommunications Conference, San Francisco, CA 2003.Google Scholar
 Qi Y, Kobayashi H: On geolocation accuracy with prior information in nonlineofsight environment. Proceedings of IEEE Vehicular Technology Conference, Vancouver, Canada 2002.Google Scholar
 Qi Y, Kobayashi H: CramerRao lower bound for geolocation in nonlineofsight environment. Proceedings of IEEE Conference on Acoustics, Speech and Signal Processing, Orlando, FL 2002.Google Scholar
 Gustafsson F: Statistical Sensor Fusion. Studentlitteratur, Lund; 2010.Google Scholar
 Qi Y, Kobayashi H: Adaptive Filtering and Change Detection. Wiley, Ltd; 2001.Google Scholar
 Parsons J: The Mobile Radio Transmission Channel. 2nd edition. Wiley, Ltd., Chichester, England; 2000.View ArticleGoogle Scholar
 Bjork A: Numerical Methods for Least Squares Problems. SIAM, Philadelphia, USA; 1996.View ArticleGoogle Scholar
 Ljung L: System Identification, Theory for the User. 2nd edition. Englewood Cliffs, NJ; 1999.Google Scholar
 Dennis J Jr, Schnabel B: Numerical Methods for Unconstrained Optimization and nonlinear Equations. PrenticeHall, Englewood Cliffs, NJ; 1983. ser. PrenticeHall series in computational mathematicsGoogle Scholar
 Carson Y, Maria A: Simulation optimization: methods and applications. Proceedings of the Winter Simulation Conference, Atlanta, Georgia 1997.Google Scholar
 Kay S: Fundamentals of signal processingestimation theory. Prentice Hall; 1993.MATHGoogle Scholar
 Lehmann E: Theory of Point Estimation. Wadsworth & Brooks/Cole; 1991. ser. Statistical/Probability seriesView ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.