利用条件概率和Gibbs抽样技术为分布估计算法构造通用概率模型

本文针对传统分布估计算法在建立概率模型时面临的各种困难,提出一种基于条件概率和Gibbs抽样的概率模型,能有效改进分布估计算法的通用性.使用该模型的分布估计算法利用进化过程中有前途的优秀个体构造出多个监督学习样本集,并对每个样本集估计出对应分量的条件概率,再使用这一组条件概率进行Gibbs抽样产生新的个体替代种群中的劣等个体.通过仿真实验表明,改进后的算法能够求解出可加性降解函数的全局最优解,表现出较强的全局优化能力.

General stochastic model for algorithm of distribution estimation with conditional probabilities and Gibbs sampling

A stochastic model based on conditional probability and Gibbs sampling is proposed to cope with the modeling problems occurred in traditional algorithms for distribution estimation, and extends the generality of the algorithm. The algorithm with this model takes promised individuals in the evolution process to form supervised training sets. For each of such sets, we estimate the conditional probability of a component given other components, and execute a Gibbs sampling procedure to generate new candidates for replacing inferior ones. The result of computer experiments shows that the improved algorithm can obtain the global optimum of additively decomposed functions, demonstrating a strong ability in global optimization.