Simplex

In Fig. 7.2, we show now the situation for dimensions till three. We denote the number of vertices for the simplex in the dimension d as n d, 1), the number of edges as n(d, 2), the number of faces as n d, 2), etc. Then we observe the following scheme as given in Table 7.2. The scheme in Table 7.2 is looking like a Pascal triangle, but in the column nil the 1 is missing. In fact, from the construction of the simplex from dimension zero upward the rule for the number of elements is 

The coefficients of x are the number of vertices p) in the simplex of dimension d, the coefficients of x are the number of edges (k), etc. On the other hand 

This holds at least for the simplex. Note that there is only one element of the highest dimension under consideration, e.g., one face in the triangle and one volume in the tetrahedron. 

Note that in Eq. (7.11) the point is associated with the edge with the plane with x and the volume with x . However, in common opinion the point is associated with zeroth dimension, the edge with the first dimension, etc. 

Into the simplex, we can introduce some vertices, edges, and surfaces by certain geometrical operations, as exemplified in Fig. 7.3. 

A number of direct ways for linking atomistic and meso-scale melt simulations have been proposed more recently. The idea behind these direct methods is to reproduce structure or thermodynamics of the atomistic simulation on the meso-scale self-consistently. As this approach is an optimization problem, mathematical optimization techniques are applicable. One of the most robust (but not very efficient) multidimensional optimizers is the simplex optimizer, which has the advantage of not needing derivatives, which are difficult to obtain in the simulation. The simplex method was first applied to optimizing atomistic simulation models to experimental data. We can formally write any observable, like, for example, the density p, as a function of the parameters of the simulation model Bj. In Eq. [2], the density is a function of the Lennard-Jones parameters. 

This mathematical identification means that we interpret oiu simulation along with the subsequent analysis of the observable by evaluating a complex function. This function, in multidimensional space, can be optimized as can any mathematical function. For an optimizer to be applicable, one must define a single-valued function with a minimum (or maximiun) at the desired target as, for example, the sum of square deviations from target values in Eq. [3]. 

Every function evaluation requires a complete equilibration sequence (either molecular dynamics (MD) or Monte Carlo (MC)) for the given parameters, followed by a production run and the subsequent analysis. To ensure equilibration, one must be certain that no drift in the observables exists, for which an automatic detection of equilibration was developed. It has been shown that derivatives of observables with respect to simulation parameters can be calculated in some cases, paving the way for more efficient optimizers.  

A drawback of the simplex and other analytical optimizers is the unavailability of numerical potentials. What is needed is a relatively small set of parameters, B , defining the entire parameter space. The limit is typically 4-6 independent parameters, because any additional dimension increases the need for computing resources tremendously. A typical choice for such parameters is a Lennard-Jones-like expansion in Eq. [5] 

Care must be taken that the same vertex is not considered as inferior in consecutive iterations. 

FIGURE 24. Reflexion of the vertex w with the worst response through 

In the modified simplex method C425 l further steps are necessary to compute an optimum new vertex and to find the maximum response with maximum speed and efficiency  

If the reflected vertex w. is better than the best vertex w. of the 

If has a better response than then la u sed as the new vertex otherwise, w is the new vertex. 

---

A systematic comparison of two sets of data requires a numerical evaluation of their likeliness. TOF-SARS and SARIS produce one- and two-dhnensional data plots, respectively. Comparison of sunulated and experimental data is accomplished by calculating a one- or two-dimensional reliability (R) factor [33], respectively, based on the R-factors developed for FEED [34]. The R-factor between tire experimental and simulated data is minimized by means of a multiparameter simplex method [33]. 

Spendley W, Next G R and Himsworth F R 1962 Sequential application of simplex designs in optimization and evolutionary operation Technometrics 4 441... 

Muller K and Brown L D 1979 Location of saddle points and minimum energy paths by a constrained simplex optimization procedure Theor. Chim. Acta 53 75... 

Multichannel time-resolved spectral data are best analysed in a global fashion using nonlinear least squares algoritlims, e.g., a simplex search, to fit multiple first order processes to all wavelengtli data simultaneously. The goal in tliis case is to find tire time-dependent spectral contributions of all reactant, intennediate and final product species present. In matrix fonn tliis is A(X, t) = BC, where A is tire data matrix, rows indexed by wavelengtli and columns by time, B contains spectra as columns and C contains time-dependent concentrations of all species arranged in rows. 

Fig. 5.4 The three basic moves permitted to the simplex algorithm (reflection, and its close relation reflect-and-expmd contract in one dimension and contract around the lowest point). (Figure adapted from Press W H, B P Flannery,...

Let us consider the application of the simplex method to our quadratic function,/ = + 2y ... 

Fig. 5.5 The first few steps of the simplex algorithm with the function + 2i/. The initial simplex corresponds to the triangle 123. Point 2 has the largest value of the function and the next simplex is the triangle 134. The simplex for tire third step is 145.

Simplex optimizations have been tried in the past. These do not assume a quadratic surface, but require far more computer time and thus are seldom incorporated in commercial software. Due to the unavailability of this method to most researchers, it will not be discussed further here. 

Lieb, S. G. Simplex Method of Nonlinear Least-Squares—A Logical Complementary Method to Linear Least-Squares Analysis ofData, /. Chem. Educ. 1997, 74, 1008-1011. 

The initial simplex is determined by choosing a starting point on the response surface and selecting step sizes for each factor. Ideally the step sizes for each factor should produce an approximately equal change in the response. For two factors a convenient set of factor levels is (a, b), a + s, h), and (a + 0.5sa, h + 0.87sb), where sa and sb are the step sizes for factors A and B. Optimization is achieved using the following set of rules ... 

Rule 1. Rank the response for each vertex of the simplex from best to worst. 

Rule 3. If the new vertex has the worst response, then reject the vertex with the second-worst response, and calculate the new vertex using rule 2. This rule ensures that the simplex does not return to the previous simplex. 

Because the size of the simplex remains constant during the search, this algorithm is called a fixed-sized simplex optimization. Example 14.1 illustrates the application of these rules. 

Find the optimum response for the response surface in Figure 14.7 using the fixed-sized simplex searching algorithm. Use (0, 0) for the initial factor levels, and set the step size for each factor to 1.0. 

The resulting simplex now consists of the following vertices Vertex Factor A Factor B Response... 

The calculation of the remaining vertices is left as an exercise. The progress of the completed optimization is shown in Table 14.3 and in Figure 14.10. The optimum response of (3, 7) first appears in the twenty-fourth simplex, but a total of 29 steps is needed to verify that the optimum has been found. 

Progress of Fixed-Sized Simplex Optimization for Response Surface in Figure 14.10... 

Progress of a fixed-sized simplex optimization for the response surface of Example 14.1. The optimum response at (3, 7) corresponds to vertex 25. 

The following set of experiments provides practical examples of the optimization of experimental conditions. Examples include simplex optimization, factorial designs used to develop empirical models of response surfaces, and the fitting of experimental data to theoretical models of the response surface. 

A variable-size simplex optimization of a gas chromatographic separation using oven temperature and carrier gas flow rate as factors is described in this experiment. 

Sangsila, S. Labinaz, G. Poland, J. S. et al. An Experiment on Sequential Simplex Optimization of an Atomic Absorption Analysis Procedure, /. Chem. Educ. 1989, 66, 351-353. 

This experiment describes a fixed-size simplex optimization of a system involving four factors. The goal of the optimization is to maximize the absorbance of As by hydride generation atomic absorption spectroscopy using the concentration of HCl, the N2 flow rate, the mass of NaBH4, and reaction time as factors. 

---