面向调控网络参数学习的无迹粒子滤波算法

目前基于微分方程模型学习网络参数的工作普遍基于卡尔曼滤波器，对所分析系统有线性假设前提，而基因调控网络具有强非线性，因此需要更适用于非线性模型的方法。提出了一种基于无迹粒子滤波器学习基因调控网络参数的方法，由于粒子滤波方法不受模型线性假设的约束，因此能够对非线性系统进行更好的拟合。通过对Repressillar模型中隐变量与未知参数的估计并与无迹卡尔曼滤波器所获结果的比较，提出的算法有效减少了估计误差。并对粒子数目对结果的影响进行了分析。相较于卡尔曼滤波器，无迹粒子滤波方法对于调控网络参数学习精度更高。粒子数目太少或太多都会减弱估计精度，因此选择适当的粒子数目非常重要。

Estimating parameters in gene regulatory network oriented unscented particle filter

The recent researches on estimation of parameters on Gene Regulatory Networks（GRN） by differential equations are generally based on Kalman Filtering Model（KFM）.It makes assumptions that the system analyzed is linear.However, GRN is obviously non-linear system,so great deviation error will happen.Here a method is presented to estimate the parameters and hidden variables of GRN based on Unscented Particle Filtering（UPF）.It makes better fitness than KFM due to free of the premise that the model is linear.By comparison of the estimation of the hidden variables and parameters of Repressilator between UPF and Unscented Kalman Filter（UKF）,advantage of this method on reduction of estimation error is validated. The effect of the amount of particles on result is simultaneously analyzed.UPF is more accurate than UKF in estimating the parameters of GRN.Deficiency or overabundance particles both will weaken accuracy of estimation,so selection on the quantity of particles is significant.