Page - 82 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3. ModelingMicrowaveHeating trainingsetDas inclassicalbatch learning, thebasic idea is thatat eachtimek, theformerNsdatapairs(slidingwindow)areusedfor thetraining(fromtimek−Ns+1 to thecurrent timek). Therefore thecost function ischangedinto JSW(k) = 1 2Ns k∑ q=k−Ns+1 ( Y(q)−YNN(q) )T( Y(q)−YNN(q) ) . (3.89) Ateachtimek,notonlythecurrentobserveddatapairbutalsoNs former data pairs are included in this cost function JSW(k). The similar idea is widely used in neural network based time series predictions, suchas in [FDH01]. These three algorithms are all fast and efficient online NN training algorithms that are implemented in practical applications. They have been tested and compared in real experiments of HEPHAISTOS, and correspondingresultscanbefoundinchapter 5. Furtherdetailsabout the neural network design and implementations refer to [HDB+96], [Roj96], [Hay98]and[AN15]. 82