Page - 72 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

3. ModelingMicrowaveHeating weights accordingly. The learning process is based on the goal of the network, normally a objective or cost function. Unsupervised learn- ing is more used in applications of exploring statistical structure of theoverall inputspace,suchaspatternclassification,decisionmaking or data compression. In control related fields, unsupervised learning canbealsoappliedinthecontrollerdesign,wherethecontrollingrule is discovered automatically. Unsupervised learning will be explained in thechapter 4withmoredetails. Reinforcement learning Comparedwithsupervisedandunsupervisedlearning,reinforcement learning is a modified combination of both [Bar98]. In reinforcement learning, for each input UNN there is also no desired output Yd. In- stead, here the network receives a feedback signalR from the plant (system being controlled) [Bar98], called reward or cost. The reward (cost) is not the direct output of the real controlled plant, but a real number indicating how good (bad) the corresponding input UNN is. Thefinalobjectiveofreinforcement learningis tochooseaninputpol- icypi tomaximize theoverall rewards. Reinforcement learning is essentially closer to the principle of hu- man behaviors, which is based on the concept of learning from ex- periences and interactions with the environment. Compared with su- pervised learning, reinforcement learning is much more flexible. In theory, an optimal control rule can be obtained by using a trial-and- learn strategy [Bar98], and no concrete system model is needed in advance. More important, reinforcement learning is suitable for con- trolling more complicated tasks which are difficult to be described by middle-complexity mathematical models. Due to these advantages, reinforcement learning method is also used in the temperature con- trol system of HEPHAISTOS, and detailed information will be given inchapter 4. 72