Page - 71 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3.3. Black-boxModeling of the network meets the learning requirement. Compared with the first fixed-topology scheme, the second learning strategy needs more computation resources and time because of the involvement of the topology adjustment [SM96]. For applications requiring fast or on- line learning (explained in the later section 3.3.2), the second strategy isnotsuitable. Therefore inthisdissertation, thefixed-topologylearn- ing scheme is used. For this scheme, there are mainly three different learningapproachesduetodifferent tasksandapplications. Supervisedlearning Neural Network UNN ∑ YNN - + Yd e Figure3.7. Principleofsupervised learning The name of supervised learning indicates the idea of involving a su- pervisorthatteachesthenetworktolearn. Theprincipleofsupervised learningisshowninfigure 3.7 [HDB+96]. Insupervisedlearning, the vectorUNN istheinputoftheNNandYd isthecorrespondingdesired (correct) output vector. They are given as data pairs ( UNN,Yd ) . The objectiveofsupervisedlearningistofindtheoptimal(correct)weights forall links thatmaketheoutputof theNNYNNequivalent to thede- sired output Yd for all trained data pairs. Supervised learning is the most common learning method used in the system modeling and the system identification. Detailed introductions of supervised learning willgiven later in the followingsection 3.3.2. Unsupervisedlearning Unlikesupervisedlearning, there isnoexplicit targetoutputorsuper- visor regarding each input in unsupervised learning [HDB+96]. In- stead of the desired output Yd, for each input vector UNN, the net- work itself decides what is the corresponding output and adjust its 71