Page - 181 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

5.3. ResultsofDifferentControlMethods After the first trial, the preliminarily trainedQ(λ) is directly used in the second trial, just as the case in the NNC experiments. The first raising-temperature period of the second trial is controlled using the nonlinear MPCmethod. In the flat-temperature period, the minimum temperature isoscillatingaroundthetarget temperature, thereforeac- cording to the principle defined in the hybrid control scheme (figure 4.14), the activated controller switching between the ’no power’ con- trol and theQ(λ) controller. The temperature state mainly stays as (Smax = 0,Smin = 1), which means the minimum temperature is in the desired temperature range but the maximum temperature is not. Although the locations states may vary during this oscillation period, all state-action values related with (Smax = 0,Smin = 1) are tested for multipletimesandupdatedtoahigheraccuracythantheothers. After theoscillation, thestatesofthesystementersintothedesiredtempera- turerangewith(Smax= 1,Smin= 1),whichmeansall temperaturesin the controlled area are within the desired range and the control target isachieved. InordertofurthertesttheeffectivenessoftheQ(λ)controller, thethird trial iscarriedoutconsecutively. Thestate-actionvaluestrainedinthe second trial is loaded in the third trial. The raising-temperature pe- riod of this trial is controlled by a NN controller. On the one hand, the results in figure 5.37c are similar with the results in the second trial. Both the maximum and the minimum temperatures stay in the desired range for the most time of this trial, which means theQ(λ) controller is effective to control the temperature distribution. On the other hand, its effectiveness can not be fully confirmed by the results in figure 5.37c. Because the temperature state in the third trial is the same as the state in the second trial. In other words, after the update process in the second trial, theQ(λ) controller may already have ac- curate state-action values for all state-action pairs related to the state (Smax = 0,Smin = 1). But it does not mean thisQ(λ) controller also have the capability to deal with other temperature states. For exam- ple, if the temperature state in the third trial is (Smax = 1,Smin = 0) (themaximumtemperatureiswithintherangebuttheminimumtem- perature is not), the controller may need a long time to update the state-actionvaluesrelatedwith this stateagain. 181