A Systematic Approach To Modeling And Optimization In The Multi-agent Combination For Controlling Bipolaris Maydis

Southerncornleafblight(SCLB)causedbytheplantpathogenBipolarismaydisisone of the major diseases afecting the cultivation and yield of maize and is widespreadthroughout most corn-growing areas. This disease can cause devastating yield lossesof up to68%every year. For controlling SCLB, there exists pathogen resistance andnegative environmental impact using single fungicide. Employing fungicides combi-nations should be overcome these disadvantages owing to their possible synergistic ef-fects. However, the number of possible combinations increases exponentially with theincrease of the number of agents and their concentrations. It is extremely challengingto identify efective agent combinations by trial and error from all possible combina-tions. In this thesis, we report on the following two strategies to handle this problem:data-driven modeling and stochastic optimization algorithms. The thesis consists offour main parts:1. In this part, we articulate the disadvantages of using single agent against SCLB.A combinatorial approach is proposed.2. Based on the system, in which three fungicides are used to inhibit the growth ofBipolaris maydis, using130data points (810in total) two data driven models areconstructedbyrespectivelyemployingneuralnetworkandsupportvectorregres-sion(SVR).Theanalysisofthepredictionabilityofthemodelsdemonstratesthatthe model based on the SVR is better to the construction of the response model.Using the predictive model and cluster analysis, the combinations are optimized.Additional experiments show that the optimal combinations can reduce the totalfungicide amount by2/3and achieve an inhibition rate higher than90%. 3. On the basis of diferential evolution algorithm, we present an approach to op-timizing agent combinations by the iteration of experiments. The simulationresults demonstrate that using this method the optimal combinations can be ob-tained using about150test data points.4. Thispartincludesasummaryofthetwoapproachesandproposesfutureresearchtopics. The proposed approaches can be generalized to handle the very similarcombinatorial problems.