A Traffic Forecasting Modle Using Adaptive BMO Algorithm Trained Neural Network

. Aiming at the short-comings of existing short-term traffic flow forecasting methods, such as low precision, Model is too complex and large computational cost, an adaptive BMO algorithm optimized BP neural network （ BMOA-NN ） algorithm is proposed. The algorithm improves the traditional BMO algorithm by introducing adaptive mechanism, The dynamic adjustment of the grouping parameters of the algorithm according to the variance of fitness, make the balance between local search and global search ability; Modeling the short-term traffic flow problem, The simulation model of the algorithm is realized, and the precision of the calculation is improved and the complexity of the calculation is reduced. The experimental results show that the proposed model can reduce the prediction error by 18.2% ~ 23.7% in the same time. ,


Introduction
In recent years, the intelligent transportation system has made great progress, but the short-term traffic flow forecast as a bottleneck of urban transportation network is characterized by high dynamic, uncertainty, non-linearity, cyclicality and spatial correlation. Need to continue to explore more accurate predictive models For the short-term traffic flow forecasting, domestic and foreign experts and scholars have adopted the linear theory method (historical average model [1] , time series model [2] , state space model [3] ); nonlinear theory method (neural network prediction model [4] , nonparametric regression prediction model [5] , support vector machine regression prediction model); mixed theory method (KARIMA [7] , ATHENA [8] , unit neural network model [9] , fuzzy neural network model [10] ) and simulation model theory. In the above method, the prediction of short-term traffic flow is not adaptive and the robustness is low. Some scholars have used the intelligent algorithm to optimize the BP (Back Propagation) neural network [11] , and the short-term traffic flow prediction model (genetic neural network Model [12] , improved particle swarm optimization algorithm based on BP neural network model [13] ), with the help of the global search capability of intelligent algorithms and the feedback mechanism of BP neural network, the prediction model has good real-time and self-The intelligent algorithm adopted by the model lacks the concept of grouping, can not achieve the balance between global search and local optimization, resulting in the

(7)
Where tpi is the desired output value; Opi is the actual output value. (4) self-learning model The self-learning process of BP neural network is to automatically correct the process of connecting the weight matrix Wij between the lower node and the upper node. The self-learning model is expressed as:

BMO Algorithm
The BMO algorithm originated from the observation of the use of high quality gene reproduction in birds and has been widely used in many fields [16]. The basic method is to abstract the problem so that it can be represented by a code to represent a solution to the problem. The algorithm first randomly generates a certain number of codes to represent the initial population, and classifies all the codes into male and female according to the degree of fitness of the fitness function, where the female is a better coding. The algorithm is divided into four types of iterations: one-sex breeding, monogamy, polygamy and promiscuity. Each code is treated as a vector, with to represent the male, to represent the female, is the new individual, ω is the weight factor with time, is the 1 × n vector, each component evaluates to a random number in [0,1], n is the dimension of the problem, mcf is the mutation control factor, is the random number in [0,1], u and l is the upper and lower bounds of the problem.
(1) sexual reproduction can be expressed as: if ： then (2) (3) In true nature, polygamy is usually a male and many females mating breeding to produce a group of offspring individuals, but in the BMO algorithm for the convenience of research, we believe that such a male and more females Mating only produces a single offspring. So polygamily expressed as: then (13) (4) Promiscuous refers to the male and female has no fixed husband and wife relationship, so the model has a great random, so in the BMO algorithm, we directly through the coding rules to re-generate some of the individuals.

Improved BMO algorithm
In order to better solve the practical problems, many intelligent algorithms have been improved to the adaptive algorithm [17] to optimize the performance of the algorithm. Since the BMO algorithm has four breeding types, each type has a parameter to represent the proportion of the population of that type, and a control factor to control the probability of occurrence of propagation. We first analyze the four types of functions in the algorithm, and then discuss how to set the adaptive function. First of all, the entry into the Sexual reproduction group is the best individual in the whole population, each individual is not affected by other individuals, refused to obtain diversity from other individuals, to keep the gene not to be degraded, so unisexual Breeding has an elite retention effect. Monogamous grouping is the most common way of breeding birds, because this breeding method can make the individual to obtain a certain diversity, but also to ensure the stability of their own genes, it is BMO algorithm stability protection. Polygamous grouping to adopt multi-partner breeding methods, in the algorithm, progeny individuals obtain genes from multiple parent individuals, so that their descendants can obtain great diversity. However, this breeding method is not stable, may produce poor individuals, but this group can avoid the algorithm into a precocious. In the algorithm, we usually remove the individuals in the abortion group directly, and then re-randomly generate the same number of individuals.
Control factor: Where f is the evaluation function of the BMO algorithm, Max is the maximum fitness value for all individuals in the nth generation population, and the initial value of mcf is 1.
We use the fitness variance to dynamically determine the proportion of the population of each group. Assuming that the number of individuals of single-sex reproduction and monogamy group is n, fi is the fitness value of the i-th individual, favg is the average fitness value of these individuals, is the fitness variance of these individuals, can be defined as: Where μ is the normalized scaling factor, its function is to limit the size of , μ can take any value, but it is necessary to ensure that the absolute value of is not more than 1, and μ is also the same as the iterative process Gradually change the value of μ can be used as follows: The variance of the fitness value reflects the degree of convergence of the two groups. The smaller the , the more the convergence is, in this case, the number of individuals in the group of parthenogenetic groups and monogamous groups should be reduced, and the number of individuals in polygamous group and overdue groups should be increased, The Polygamous group: Where monogamo initial ratio is similar t The pro Where α

Conclusion
In this paper, an adaptive parameter improvement based on fitness variance is proposed for the parameter setting of BMO algorithm, which makes the improved BMO algorithm have better performance. The improved BMO algorithm is applied to the training of BP neural networks for short -term traffic flow prediction, and the BMOA -NN model is proposed. After the simulation experiment, the prediction results are evaluated by a variety of error evaluation criteria. It is found that the prediction results of the BMOA -NN model are more accurate and robust than the other three models, and can provide more accurate and accurate prediction for traffic guidance and traffic control.