改进的GA-BP神经网络模型在财产犯罪预测中的应用

Property Crime Forecast Based on Improved GA-BP Neural Network Model

  • 摘要: 发现犯罪时空分布规律并预测犯罪发生,是提高警务策略有效预防、控制犯罪的重要方法。在分析财产犯罪时空规律的基础上,利用BP神经网络模型自动学习训练各因子与财产犯罪的非线性关系,建立了财产犯罪预测模型。针对BP神经网络模型易陷入局部最优和模型不稳定的缺陷,提出了利用遗传算法(GA)选择各因子最优的初始化权重和参数,并以此作为BP神经网络模型的初始化权重矩阵,通过对历史数据的学习及训练建立了改进后的GA-BP神经网络模型。利用某市2007~2012年财产犯罪、人口、GDP、土地利用等35个综合影响因子数据,对改进前后的模型进行了预测对比试验。结果表明,改进后的GA-BP神经网络模型成功克服了BP模型的缺陷,收敛迭代最小次数从117次改进到8次;10次计算收敛迭代次数最大误差从370次提高到5次;模型预测精度(RMES)从0.043 0提高到0.019 95。

     

    Abstract: To discover the spatial-temporal distribution and estimate the occurrence of crimes is an important method of improving policing strategies and preventing and controlling crimes effectively. In this paper, a prediction model for property crimes is first established based on the analysis of the spatial-temporal distribution of property crimes by using the BP(back propagation) neural network to train and learn the non-linear relationship between factors and crimes automatically. Aiming at the defects of BP neural network model of easily trapping in local optimum and instability, an improved GA-BP neural network model is then put forward, which uses the genetic algorithm (GA) to select the optimal initial weights and parameters for BP neural network model so as to learn and train the historical data. Finally, in order to evaluate whether our improved GA-BP neural network model is better than the BP neural network model in forecasting property crimes, a comparative experiment between those two predictive models is carried out with the data of 35 comprehensive impact factors from 2007 to 2012, such as property crimes, population, GDP, land utilization and so on. According to the study results, the improved GA-BP neural network model overcomes the defects of BP model successfully and shows a better performance in predicting property crimes. On the one hand, the minimum number of convergent iteration is reduced from 117 to 8. On the other hand, the maximum error of the ten times of calculation of the iterations is reduced from 370 to 5. Additionally, the prediction precision RMES is improved from 0.043 0 to 0.019 95.

     

/

返回文章
返回