Research on the Improvement of Efficiency of Edas for Optimization

Estimation of Distribution Algorithm (EDA) is a new concept in the field of evolutionary computation, which is motivated by the idea of building a probabilistic model of the population to prevent the disruption of building blocks in subsequent generations. It throws out the genetics of evolutionary computation; instead, produces solutions by using techniques developed in statistics, artificial intelligence, and clustering. Many researches have been carried out in this field and different variants of EDAs have been proposed, but very few of them are efficient for harder problem instances. The aim of this work is to develop an EDA that can solve complex problems more efficiently and more reliably. Reinforcement Learning Estimation of Distribution Algorithm (RELEDA), a univariate EDA in binary search space, has been proposed in this thesis. It provides better quality of solutions to problems with many local minima, more efficiently and more reliably than other existing EDAs by combining techniques from EDA and reinforcement learning. Problems with continuous variables can be solved by the algorithm if they are binary-encoded, but the precisions may be lost. To use real-coded variables of real-world problems another algorithm named Real-Coded Estimation of Distribution (RECEDA), which is a multivariate EDA in continuous domain and generates offspring based on the Cholesky decomposition of the covariance matrix of the variables, has been suggested. RECEDA have been tested on some bench-mark problems of function-optimization and it outperforms other EDAs in continuous domain.