13.7 Neural-Network-Based Optimization 727 subject to _ µ f (X) _ µg(l) j ,) (X _ µg(u) j 13.6.4 (X ,) j = 1, 2 . ., . , m j = 1, 2 . ., . , m (13.59) Numerical Results The minimization of the error between the generated and specified outputs of the four-bar mechanism shown in Fig. 13.9 is considered. The design vector is taken as X = {a b c ffi , T} . The mechanism is constrained to be a crank-rocker mechanism so that a Š b _ 0, a Š c _ 0, a_1 d = [(a + c) Š (b + 1)][(c Š a) 2 Š (b Š 1 2) ]_0 The maximum deviation of the transmission angle (µ) from 90 ff is restricted to be less ff than a specified value, tmax = 35 . The specified output angle is . • s(_) = ff 20 + •3 , 0 unspecified, 240 ff _ _ _ 240 ff ff _ _ < 360ff Linear membership functions are assumed for the response characteristics [13.22]. The optimum solution is found to be X = {0.2537 0 .8901 0 .8865 Š 0 .7858 with f _ = 1.6562 and _ = 0.4681. This indicates that the maximum level of satisfaction that can be achieved in the presence of fuzziness in the problem is 0.4681. The transmission angle constraint is found to be active at the optimum solution [13.22]. 13.7 NEURAL-NETWORK-BASED OPTIMIZATION The immense computational power of nervous system to solve perceptional problems in the presence of massive amount of sensory data has been associated with its parallel r3 = b r4 = c q3 r2 = a q4 w2 q2 = f b 1 Figure 13.9 Four-bar function generating mechanism. Š10 T.} 728 Modern Methods of Optimization processing capability. The neural computing strategies have been adopted to solve optimization problems in recent years [13.23, 13.24]. A neural network is a massively parallel network of interconnected simple processors (neurons) in which each neuron accepts a set of inputs from other neurons and computes an output that is propagated to the output nodes. Thus a neural network can be described in terms of the individual neurons, the network connectivity, the weights associated with the interconnections between neurons, and the activation function of each neuron. The network maps an input vector from one space to another. The mapping is not specified but is learned. Consider a single neuron as shown in Fig. 13.10. The neuron receives a set of n inputs, x i, i = 1, 2, . . . , n, from its neighboring neurons and a bias whose value is equal to 1. Each input has a weight (gain) w i associated with it. The weighted sum of the inputs determines the state or activity of a neuron, and is given by a = _ i +1n i i=1 w x 2 T n T = W X, where X = {x 1x · · · x 1} . A simple function is now used to provide a mapping from the n-dimensional space of inputs into a one-dimensional space of the output, which the neuron sends to its neighbors. The output of a neuron is a function of its state and can be denoted as f (a). Usually, no output will be produced unless the activation level of the node exceeds a threshold value. The output of a neuron is commonly described by a sigmoid function as 1 (13.60) f (a) = 1 + e Ša which is shown graphically in Fig. 13.10. The sigmoid function can handle large as well as small input signals. The slope of the function f (a) represents the available gain. Since the output of the neuron depends only on its inputs and the threshold value, each neuron can be considered as a separate processor operating in parallel with other neurons. The learning process consists of determining values for the weights w i that lead to an optimal association of the inputs and outputs of the neural network. x1 xn xn+1 = 1 w1 wn a f (a) wn+1 (Bias) f (a) 1.0 0.5 a 0 Figure 13.10 Single neuron and its output. [12.23], reprinted with permission of Gordon & Breach Science Publishers. 13.7 Neural-Network-Based Optimization 729 Several neural network architectures, such as the Hopfield and Kohonen networks, have been proposed to reflect the basic characteristics of a single neuron. These architectures differ one from the other in terms of the number of neurons in the network, the nature of the threshold functions, the connectivities of the various neurons, and the learning procedures. A typical architecture, known as the multilayer feedforward network, is shown in Fig. 13.11. In this figure the arcs represent the unidirectional feedforward communication links between the neurons. A weight or gain associated with each of these connections controls the output passing through a connection. The weight can be positive or negative, depending on the excitatory or inhibitory nature of the particular neuron. The strengths of the various interconnections (weights) act as repositories for knowledge representation contained in the network. The network is trained by minimizing the mean-squared error between the actual output of the output layer and the target output for all the input patterns. The error is minimized by adjusting the weights associated with various interconnections. A number of learning schemes, including a variation of the steepest descent method, have been used in the literature. These schemes govern how the weights are to be varied to minimize the error at the output nodes. For illustration, consider the network shown in Fig. 13.12. This network is to be trained to map the angular displacement and angular velocity relationships, transmission angle, and the mechanical advantage of a four-bar function-generating mechanism (Fig. 13.9). The inputs to the five neurons in the input layer include the three link lengths of the mechanism (r 2, r 3, and r 4) and the angular displacement and velocities of the input link (_ 2 and - 2). The outputs of the six neurons in the output layer include the angular positions and velocities of the coupler and the output links (_ 3, - 3, _ 4, and - 4), the transmission angle (_ ), and the mechanical Outputs Output layer Hidden layer Input layer Inputs Figure 13.11 Multilayer feedforward network. [13.23], reprinted with permission of Gordon and Breach Science Publishers. 730 Modern Methods of Optimization q3 q4 r2 w3 r3 w4 r4 g q2 h w2 Figure 13.12 Network used to train relationships for a four-bar mechanism. [12.23], reprinted with permission of Gordon & Breach Science Publishers. advantage (.) of the mechanism. The network is trained by inputting several possible combinations of the values of r 2, r 3, r 4, _ 2, and - 2 and supplying the corresponding values of _ 3, _ 4, - 3, - 4, _ , and .. The difference between the values predicted by the network and the actual output is used to adjust the various interconnection weights such that the mean-squared error at the output nodes is minimized. Once trained, the network provides a rapid and efficient scheme that maps the input into the desired output of the four-bar mechanism. It is to be noted that the explicit equations relating r2 , r 3, r 4, _ 2, and - 2 and the output quantities _ 3, _ 4, - 3, - 4, _ , and . have not been programmed into the network; rather, the network learns these relationships during the training process by adjusting the weights associated with the various interconnections. The same approach can be used for other mechanical and structural analyses that might require a finite-element-based computations. truss described in Section 7.22.1 (Fig. of 7.21)thewasstructural considered with of constraints on Numerical Results. The minimization weight the three-bar the cross-sectional areas and stresses in the members. Two load conditions were considered with P = 20,000 lb, E = 10 × 10 6 psi, = 0.1 lb/in 3, H = 100 in., fimin = (l) (u) 2 2 = 0 1. in (i = 1, 2), and Ai = 5 0. in (i = 1, 2). Š15,000 psi, fimax = 20,000 psi, A i optimization The solution obtained using neural-network-based is [12.23]: 0 .4079 in and f _ = 26.3716 lb. be compared 0 .788 in can This x1 x2 =_2 , 2 2, solution given by nonlinear with the programming: x 1_ = 0.7745 in 2, x _ = 0.4499 in , 2 =_ f _ = 26.4051 lb. and REFERENCES AND BIBLIOGRAPHY 13.1 J. H. Holland, Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann Arbor, MI, 1975. 13.2 I. Rechenberg, Cybernetic Solution Path of an Experimental Problem, Library Translation 1122, Royal Aircraft Establishment, Farnborough, Hampshire, UK, 1965. 13.3 D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Linearning, Addison-Wesley, Reading, MA, 1989.