Multi-Objective Evolutionary Algorithms for Ecological Process Models


Fitting an ecological process model to a set of data is frequently done by minimizing the residual sum of squares (RSS) between data and model output. However, we may need to consider many component elements of an ecological process when simulating a model rather than just one, and the RSS may not be an appropriate metric for simulation assessment. For this dissertation, a multi-objective Evolutionary Algorithm (EA) was used to fit a complex ecological process model. As an example, a model of successive hourly shoot growth of a forest tree was used. First, single-objective methods were tried with the RSS; however, the simulation results did not capture the measured data well, especially contraction periods. The current simulated growth is affected by that of the previous hours because the model includes a regression term. Thus, the fitting result could be improved if there was information about the relation between each data. The multi-objective optimization method allows us to consider contraction and extension periods separately, and these are an important phenomenon in shoot growth. Since we can set more than one objective function, each focused on particular data features. Also, if there is difficulty in achieving some criteria at the same time, analysis of differential effectiveness in capturing contraction and extension, if it occurs, could help to find what and where the deficiency of the model is. These effects of considering more than one objective function motivated using a multi-objective optimization method. Since the model is complex, many objective functions were required. I implemented elitism, a process to keep the best individuals to the next generations, but this needed to be different from that used in other EAs and obtained the following results: (1) crossover mutation rate should be determined dynamically for an efficient search; (2) from analysis of the results, deficiencies of the model were identified; (3) with the revised model, accuracy of achievement at contraction periods was improved; (4) the model reduced bias, but the error did not become small; more biological information about contraction and expansion is needed.