Page - 68 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3. ModelingMicrowaveHeating 0 4-4 0.5 1 8-8 (a)Unit step 0 4-4 0.5 1 8-8 (b)Piecewise linear 0 4-4 0.5 1 8-8 (c)Sigmoidaccordingto [HDB+96] 0 4-4 1 8-8 -1 (d)Tanh(equation 3.69) Figure3.5. Illustrationofdifferentactivationfunctions. orfeaturethat is fedfromtheinputlayer,extract themorevaluablein- formationfromtherawinformation,andthentransmitit totheoutput layer for thefinaldetermination[AN15]. In other words, the task being solved by the neural network can be decomposed into a number of small subproblems, and subproblems are solved and united layer by layer. Theoretically, the neural net- workswithmorehiddenlayers(alsoknownasdeepneuralnetworks) are able to perform more complicated and accurate approximations to highly dynamic systems, and have a better performance than the so-called shallow networks that have only one or few hidden layers. However, in deep neural networks there are a number of obstacles thatarestillnotwellunderstood,suchasthevanishinggradientprob- lem [BSF94] or influences caused by weight initializations [SMDH13]. In practice as well as many literatures [Hay98], it is suggested to use a neural network with an appropriate number of layers instead of a deepnetwork. 68