Hand processing weight calculations

The ability of an ANN to learn is its greatest asset. When, as is usually the case, we cannot determine the connection weights by hand, the neural network can do the job itself. In an iterative process, the network is shown a sample pattern, such as the X, Y coordinates of a point, and uses the pattern to calculate its output it then compares its own output with the correct output for the sample pattern, and, unless its output is perfect, makes small adjustments to the connection weights to improve its performance. The training process is shown in Figure 2.13. 

In optimization using a modular process simulator, certain restrictions apply on the choice of decision variables. For example, if the location of column feeds, draws, and heat exchangers are selected as decision variables, the rate or heat duty cannot also be selected. For an isothermal flash both the temperatures and pressure may be optimized, but for an adiabatic flash, on the other hand, the temperature is calculated in a module and only the pressure can be optimized. You also have to take care that the decision (optimization) variables in one unit are not varied by another unit. In some instances, you can make alternative specifications of the decision variables that result in the same optimal solution, but require substantially different computation time. For example, the simplest specification for a splitter would be a molar rate or ratio. A specification of the weight rate of a component in an exit flow stream from the splitter increases the computation time but yields the same solution. 

The braids were tested from — 100°C upwards at a heating rate of 1°C min 1 and a frequency of 1 Hz, using an inertial weight of 4.55 X 10 5 kg m2. The damped sine waves were processed by hand to give log decrement and modulus. The radius of the sample was calculated from the weight per unit length, as above, with the additional assumption that the sample was a right circular cylinder and that there were no voids in the sample. 

To make quantitative statements about the product internal distribution a computer program is utilized to simulate the observed excitation spectrum [10]. As input for the calculations we estimate the relative vibrational and rotational populations. Each line is weighted by the population of the initial (v, J ) level, by the Franck-Condon factor and the rotational line strength of the pump transition. At each frequency, the program convolutes the lines with the laser bandwidth and power to produce a simulated spectrum such spectra are compared visually with the observed spectra and new estimates are made for the (v ,J") populations. Iteration of this process leads to the "best fit" as shown in the lower part of Fig. 3. For this calculated spectrum all vibrational states v" = 0...35 are equally populated as is shown in the insertion. The rotation, on the other hand, is described by a Boltzmann distribution with a "temperature" of 1200 K. With such low rotational energy no band heads are formed for v" < 5 in the Av = 0 sequence and for nearly all v" in the Av = +1 sequence (near 5550 A). 

Hidden layer calculations. To update the left-hand layer of weights (between input and hidden layers), we need the information we have just calculated. Figure 4.8 shows the results of the last calculation on the network. For the sake of simplicity, the dashed nodes and weights in the network wiU be ignored. In practice, you would perform the following process on those neurons as well. The hidden node shown gave an output of 0.5 on the first pass through the network. 

---