| Data mining tasks are pivotal for the improvement of manufacturing and design processes. However, some of the hidden patterns or relationships among the data are very complex, which cannot be easily detected by traditional data mining techniques. Several example industrial applications include cell phone reliability drop testing, call failure detection, wave filter design, and simulations, etc. Gene Expression Programming (GEP) was recently developed to address this challenge in data analysis and knowledge discovery. Being an evolutionary computation method, GEP distinguishes itself by searching the global optimum through a population of candidate solutions in parallel and being able to produce solutions of any possible form with minimum requirements of pre-knowledge.; Although quite flexible, the algorithm still has limited performance with respect to complex problems since structure related information about evolving solutions is overlooked during its execution. This research aims to improve the problem solving ability of the GEP algorithm for complex data mining tasks by preserving and utilizing the self-emergence of structures during its evolutionary process.; An incremental approach has been pursued to achieve the proposed research goal, including the investigation of the constant creation methods in GEP, for identifying and promoting good solution structures; the design of a new genotype representation, namely, Prefix Gene Expression Programming (P-GEP), for establishing a solution structure preserving evolutionary process; and the introduction and implementation of self-emergent structures in P-GEP, for speeding up the learning process by reusing some evolved useful structural components and hence decomposing the complexity of the target solutions.; Benchmark testing and theoretical analysis have both demonstrated that this line of work successfully assists the evolutionary process in advocating solutions with good functional structures, and finding meaningful building blocks to hierarchically form the final solutions following a faster fitness convergence curve, especially when applied to structurally complex problems. In general, more accurate solutions, higher success rates, and more compact solution structures have been achieved compared to the original GEP algorithm and other traditional methods. |
