17 OPTIMIZATION Introduction [1]: The most general method for solving nonlinear inverse problems needs a complete exploration of the model space. Due to the fact for any other purposes other than academic, this method is usually too computer intensive, usual methods normally limit their scope to achieving some “best” model, as in a model maximizing the probability density σ M(m) or minimizing some error function S(m). If the forward problem isn’t very nonlinear, the required functions σ M(m) and/or S(m) are normally well behaved and usually have only one extreme point, which could be obtained by using the gradient method, for example, i.e., methods that use the function’s properties at a specific point mn to decide on the future search direction so as to acquire the updated model mn+1. For very nonlinear problems, however, there is a great risk that the problem won’t converge to the required extreme points but rather to local min/max points that don’t form the required solution to the problem. It has been shown that for model spaces with more than a few parameters, it is more economical to select random points in the space model, rather than building an evenly spread grid that is dense enough to guarantee that at least one point will fall in the required optimal area. Any method that uses a random (or pseudo-random) generator at any point is named Monte Carlo. The interest of Monte Carlo methods for inversion is that they can solve problems of relatively large size without any linearization. For the search for the domain of permitted models we will assume some known information of the parameter domain, which can be described in the simple form: m inf m msup ( I M ) (1) Where: M = the model space M = a particular point in M The Monte Carlo method of inversion involves using a pseudo-random number generator to generate random models within the section defined by (1). Computation of these models, m, results in the estimated data, dcal = f(m). A quantitative comparison between dcal and dobs, the observed data serves as a criterion for the suitability of any specific model, m. This computation is stopped once the number of accepted models is enough to suggest that that the model space has been explored. An interesting reference to the use of Monte Carlo methods of inversion in geophysics is the Press (1968) method of searching for the region of admissible models. Press studied the density of the Earth’s mantle, as well as the velocity of seismic waves, as a function 18 of the radius r, using as data measured eigenperiods of the Earth’s vibration, seismic wave travel time, the total Earth’s mass and the Earth’s moment of inertia. The parameters he wanted to estimate were the density ρ(r), the velocityof the longitudal waves α(r), and the velocity of the tranverse waves β(r). These functions were considered at 23 values of of r (the rest of the values were achieved by interpolation), resulting in a total of 69 parameters. Approximately five million models were randomly geneerated and tested. Of these only six gave estimated data close enough to the observed values. Figure 1 shows four of the Earth’s models obtained this way. This figure gives quite a good idea of the domain of admissible Earth models, the invers problem has basically been solved. Fig1. Four of the accepted Earth’s models of β, transverse waves velocity and ρ, the density found by Press (1968). The heavy curves define the expected region of results. The example above illustrates some of the difficulties of this method. If the region error bars are chosen too large, the number of experiments may be very large; if they are chosen too small, they will control the results. It is also difficult to determine that the number of trials suffices. The problem: This project is based on a previous study [2] done on muscle tissue. There it was shown that the tissue impedance characteristics comply with the following nonlinear mathematical equations: 19 V23( , l ,t ) K cos2 t l sin 2 1/ 2 (2) V89 ( , t , l ) K[cos 2 ( 1 ) t l sin 2 ( 1 )] 1/ 2 (3) V56 ( , t , l ) K[cos 2 ( 2 ) t l sin 2 ( 2 )] 1/ 2 (4) K = I 2a l t 1/ 2 (5) Where: σt = electic conductivity (parralel to the fibers) σl = electic conductivity perpandicular to σt (transverse to the fibers) α = angle between the direction transverse to the fibers and the straight line through the four electrodes [2] I = alternating current through the medium V = voltage response from the medium to the current I For both V and I the reference is to each channel saparetely a = distance between two adjacent electrodes. γ1,2 = represent the angles between the first row of electrodes and the second and third rows respectively The optimization concerns finding the σt, σl, α parameter values so that the above function minimization is achieved. Using the matlab function, we obtain a set of predicted parameters for different starting points. These starting points are spread randomly within a priori region based on existing literary values. By selecting the minimum of the different function’s minima achieved one can assume that the function global minimum was obtained. The tighter the ‘starting point’ lattice the higher the probability of falling on the global minimum, however, the price for that is a very complex and time consuming process. The matlab Optimization toolbox [3] offers different routines that deal with optimization of non-linear functions. These routines vary in algorithms and line search strategies. Trial and error brought us to use the fsolve matlab function. The fsolve matlab function: This function finds the roots of the non-linear equation. Values of X are found so that: F(X) = 0 20 (6) The optimization problem in this case is a non-linear least square problem and uses the Gauss-Newton or Levenberg-Marquardt methods. The choice of the algorithm and other parameters can be set manually by the user. The options used in the optimization being: 1. Terminate for x = 1e-4 (default). This defines the worst case precision required of the independent variable, x. The optimization will carry on until termination criteria is reached. Terminate for f = 1e-4 (default). This defines algorithm termination based on precision required of f at solution (In theory f = 0). Main Algorithm = Gauss Newton Search Algorithm = The line search algorithm used is a safeguarded cubic polynomial method. This method requires fewer function evaluations but rather more gradient evaluations. Max iterations = 300 (default). The default number of iterations is defined as 100*n where n is the number of independent variables. 2. 3. 4. 5. The Optimization algorithm: Choose estimation starting point Activate fsolve optimization function: Set function options different to default Sensitivity points in the algorithm: 1. Unless the function is continuous and has one minimum only, there is no guarantee that the minimum achieved is the global minimum. Starting the optimization at different starting points helps locate the global minimum. The smaller the distance between these starting points the higher the probability to arrive at the optimal results and the accuracy level rises. 2. The function to be solved should be continuous. When successful, fsolve returns one root only. fsolve may converge to a non-zero point in which case another starting point may help. 3. The noise level and characteristics affect the optimization accuracy. Sharp spikes in the function or a noise level higher than 10[dB] throw the optimization algorithm off course The Xo beginning points are limited to a defined boundary based on previous studies. Assuming these boundaries are correct and the requested estimated parameters are located within these limits this factor shouldn’t have any affect. However, if the 21 parameter values aren’t included within this boundary the estimation will not converge to any value. REFERENCES [1] Tarantola, A., ”Monte Carlo Methods”, Inverse Problem Theory Methods for Data Fitting and Model Parameter Estimation, pp 167-170, Elsvier, Netherlands,1987 [2] F.L.H Gielen; K.L. Boon, ‘Electrical Conductivity & Histological Structure of Skeletal Muscle’. [3]“Optimization Toolbox – User’s Guide”, 1992.