CN109034898A - A kind of BP neural network used car price evaluation algorithm based on improvement ant colony - Google Patents

Abstract

The invention discloses a BP neural network second-hand car price evaluation algorithm based on an ant colony algorithm optimization. A three-layer BP neural network is selected as a prototype, and an improved ant colony algorithm is used to optimize the weight initialization process of the BP neural network. Car price evaluation model, the specific steps are as follows Step 1: Data collection and preprocessing; Step 2: Determine the network topology; Step 3: Use the improved ant colony algorithm to optimize the weight initialization process of the BP neural network; Step 4: Further The optimized BP neural network is trained to predict the price of used cars. The invention provides a BP neural network second-hand car price evaluation algorithm based on ant colony optimization. Compared with the traditional algorithm, it can improve the defects of BP neural network, such as easy to fall into local optimum, slow convergence speed, and oscillation effect, so as to establish A set of online vehicle price evaluation system with practical value.

Description

A second-hand car price evaluation algorithm based on improved ant colony BP neural network

technical field

The invention relates to a second-hand car price evaluation algorithm based on an improved ant colony BP neural network, which belongs to the field of computer applications.

Background technique

Against the background of the increasing development of network marketing and Internet big data theory, combined with the growing demand of Chinese people for private cars, large-scale used car trading platforms based on neural networks have been rapidly promoted and applied. Under such a platform trading system, by analyzing and adjusting various parameters that need to be considered in the used car trading market, and establishing a suitable neural network model, the accuracy of the entire used car trading price evaluation can be improved. As big data-related technologies are widely used in the fields of machine learning and cognitive science, the BP neural network algorithm, as a calculation model mainly used to estimate and approximate functions, connects a large number of neurons for calculation, and can realize a An effective predictive model.

As a tangible asset, second-hand cars are commonly used in price evaluation methods in transactions, including replacement cost method, income present value method, current market price method and liquidation price method. Most of these existing valuation algorithms are too empirical, do not have much theoretical basis, and do not actively disclose the details of the algorithms. There are defects such as low algorithm transparency, low valuation efficiency, and low calculation accuracy. The dynamic changes of the data are reflected in the calculation process, and these are key factors in determining the price of a used car.

Contents of the invention

In order to overcome the defects of the prior art, the present invention proposes and discloses a second-hand car price evaluation algorithm based on an improved ant colony BP neural network, and utilizes the global search characteristic of the ant colony algorithm to roughly search for a certain weight range. The weight value is used as the initial weight value of the BP neural network, and then the BP algorithm is used to further optimize the network weight value, which can improve the defects in the BP neural network that are easy to fall into local optimum, slow convergence speed, and cause oscillation effects, thus establishing a set of An online vehicle price evaluation system with practical value.

In order to achieve the above object, technical solution of the present invention is achieved in that way:

An improved BP neural network second-hand car price evaluation algorithm based on ant colony algorithm. The three-layer BP neural network is selected as the prototype, and the improved ant colony algorithm is used to optimize the weight initialization process of the BP neural network, and a second-hand car price evaluation model is established. ,Specific steps are as follows:

Step 1: Data collection and preprocessing;

Step 2: Determine the network topology;

Step 3: Use the improved ant colony algorithm to optimize the weight initialization process of the BP neural network;

Step 4: Further train the optimized BP neural network to predict the price of used cars.

Preferably, the data collected in the step (1) is second-hand vehicle data and its transaction data as sample data, including vehicle serial number, production year, year and month of sale, city, displacement, new car price, mileage, vehicle usage, degree of wear and tear and personal guide price; in the step (1), the collected sample data is preprocessed: the collected sample data is normalized, specifically normalized to the interval [0,1], and the normalization formula is as follows :

X ₀ =(XX _min )/(X _max -X _min ) (1)

In formula (1), X ₀ is the normalized data, X is the original data, X _max and X _min are the maximum and minimum values of the original data set, respectively.

Preferably, the network topology in the step (2) is a three-layer BP neural network structure, including an input layer, a hidden layer, and an output layer, wherein,

Input layer: Input preprocessed sample data, each input node is mapped to a vehicle-related attribute, specifically vehicle number, production year, sales year and month, city, displacement, new car price, mileage, vehicle use, wear and tear A total of 10 neurons for degree and individual guide valence;

Hidden layer: Estimate the number of neurons in the optimal hidden layer according to the empirical formula (2),

In formula (2), H is the number of neurons in the hidden layer, I is the number of neurons in the input layer, O is the number of neurons in the output layer, and a is a constant in [1,10];

Output layer: output the actual transaction price results of used cars, and provide real-time feedback for user reference.

Preferably, the optimization process of the weight initialization in the step (3) is as follows:

(3a) Parameter setting, divide the weight interval [-5,5] of BP neural network into 100 equal parts, create a pheromone table for each weight parameter, record w _s as the sth one that needs to be optimized Weight parameter, the value range of s is [1,N], N is the total number of weights, i is the division scale value, and every two division adjacent scale values form a sub-region; τ(i) is The pheromone value corresponding to i, where the weight interval [-5,5] is evenly divided into 100 equal parts, then there are 101 divided scale values;

The total number N of the weight is determined by the number I of input layer neurons, the number H of hidden layer neurons and the number O of output layer neurons, and its calculation formula is as follows:

N=H*(I+O+1)+O (3);

Simultaneously, the initial value of setting pheromone is τ (i)=C, C ≠ 0, pheromone volatilization coefficient ρ, pheromone incremental strength Q, maximum number of iterations countmax of ACO and ACO optimization end condition ε _ACO ;

(3b) Release m ants. For any weight parameter w _s , the nth ant moves from one point to the next according to the following probability formula (4):

In formula (4), i represents the i-th division scale value of the weight parameter w _s , and the value range of i is [1,101], Indicates the sum of the pheromones of all ants in the weight parameter w _s , and j represents the jth ant. In each iteration of the neural network, the ant will update its own pheromone value according to the error. The pheromone value here is the above One of the 101 division scale values.

The nth ant passes through the division scale value of the weight parameter w _s and only once, and records the division scale value of the corresponding point. The combination of the division scale values of these points constitutes a set of weight parameters of the neural network weight parameter w _s ;

(3c) Use the second-hand car transaction data and its vehicle data as input training samples, use the weight combination obtained in step (3b) as the parameters of the neural network, and use The Sigmoid S-type activation function performs the output calculation of the neural network, as shown in formula (5):

In formula (5), net representation is hidden layer and output layer, input layer and hidden layer, the functional relation between layer and layer, and neural network activation function belongs to conventional technical means in the present invention so does not add detailed description;

After SH gets the output of the neural network, it calculates the mean square error and takes the maximum value of the mean square error, as shown in formula (6):

In formula (6), SampleNum is the number of samples, y is the expected output value, that is, the real value of the training sample, o is the actual output value of the neural network, and o is determined by formula (5), which belongs to the basic technical category of neural network;

(3d) After all ants construct the solution, record a set of weights with the smallest E, compare the minimum error E _min with ε _ACO , if E _min <ε _ACO , then directly complete the initialization process and exit, otherwise go to step (3e);

(3e) Pheromone update: The pheromone update strategy of the i-th division scale value of the weight w _s is shown in formula (7):

In formula (7), The i-th division scale value of the weight w _s corresponds to the pheromone value updated after the n-th ant in the t-generation ant colony passes through, and the value range of μ is [10,100];

(3f) Steps (3b)-(3d) are repeated until the maximum number of iteration countmax is satisfied, and the initialization process is completed.

Preferably, the process of further training the optimized BP neural network in the step (4) is: according to the ant colony algorithm, a group of weights and deviations (being the weight parameter w) that are found by the E minimum in step 3 (d) _s ) as the initial weight and deviation of the BP algorithm, calculate the error between the network output and the actual output, and propagate the error back from the output layer to the input layer, further adjust the weight and deviation, and repeat the above process until the training exit condition.

The process of further training the optimized BP neural network in step (4) of the present invention belongs to conventional technical means, so it is not described in detail.

Beneficial effects: the present invention provides an improved BP neural network second-hand car price evaluation algorithm based on the ant colony algorithm, which is introduced into the ant colony algorithm The global search ability of the algorithm is improved. Compared with prior art, the present invention has following advantage:

(1) Compared with the traditional mechanical evaluation algorithm (vehicle evaluation value = replacement cost * newness rate * adjustment coefficient), the BP neural network based on improved ant colony used car price evaluation algorithm has a higher prediction accuracy;

(2) The present invention adopts the improved ant colony algorithm to optimize the weight initialization process of the BP neural network on the basis of the original BP neural network, and improves the defects such as being easy to fall into local optimum, slow convergence speed, and oscillation effects in the BP neural network, Thus, a set of online vehicle price evaluation system with practical value is established.

(3) After the user completes the transaction, the new transaction data is put into training to maintain the high accuracy of the model, so that it can provide services to users more permanently.

Description of drawings

Fig. 1 is the data parameter table of the present invention.

Fig. 2 is a neural network structure diagram of a specific example of the present invention.

Fig. 3 is the flow chart of the BP neural network algorithm based on the improved ant colony of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the technical solutions in the application, the technical solutions in the embodiments of the application are clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of the application, and Not all examples. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

An improved BP neural network second-hand car price evaluation algorithm based on ant colony algorithm. The three-layer BP neural network is selected as the prototype, and the improved ant colony algorithm is used to optimize the weight initialization process of the BP neural network, and a second-hand car price evaluation model is established. ,Specific steps are as follows:

Step 1: Data collection and preprocessing;

Step 2: Determine the network topology;

Step 3: Use the improved ant colony algorithm to optimize the weight initialization process of the BP neural network;

Step 4: Further train the optimized BP neural network to predict the price of used cars.

Preferably, the data collected in the step (1) is second-hand vehicle data and its transaction data as sample data, including vehicle serial number, production year, year and month of sale, city, displacement, new car price, mileage, vehicle usage, degree of wear and tear and personal guide price; in the step (1), the collected sample data is preprocessed: the collected sample data is normalized, specifically normalized to the interval [0,1], and the normalization formula is as follows :

X ₀ =(XX _min )/(X _max -X _min ) (1)

In formula (1), X ₀ is the normalized data, X is the original data, X _max and X _min are the maximum and minimum values of the original data set, respectively.

Preferably, the network topology in the step (2) is a three-layer BP neural network structure, including an input layer, a hidden layer, and an output layer, wherein,

Input layer: Input preprocessed sample data, each input node is mapped to a vehicle-related attribute, specifically vehicle number, production year, sales year and month, city, displacement, new car price, mileage, vehicle use, wear and tear A total of 10 neurons for degree and individual guide valence;

Hidden layer: Estimate the number of neurons in the optimal hidden layer according to the empirical formula (2),

In formula (2), H is the number of neurons in the hidden layer, I is the number of neurons in the input layer, O is the number of neurons in the output layer, and a is a constant in [1,10];

Output layer: output the actual transaction price results of used cars, and provide real-time feedback for user reference.

Preferably, the optimization process of the weight initialization in the step (3) is as follows:

(3a) Parameter setting, divide the weight interval [-5, 5] of BP neural network into 100 equal parts, and establish a pheromone table for each weight parameter, as shown in Table 1,

Table 1 parameter information table

Note that w _s is the sth weight parameter that needs to be optimized. The value range of s is [1,N]. divided adjacent scale values constitute a sub-region; τ(i) is the pheromone value corresponding to Ai, where the weight interval [-5, 5] is evenly divided into 100 equal parts, then there are 101 divided scale values;

The total number N of the weight is determined by the number I of input layer neurons, the number H of hidden layer neurons and the number O of output layer neurons, and its calculation formula is as follows:

N=H*(I+O+1)+O (3);

At the same time, set the initial value of the pheromone to τ ₀ , the pheromone volatilization coefficient ρ, the pheromone incremental strength Q, the maximum iteration countmax of ACO and the ACO optimization end condition ε _ACO ;

(3b) Release m ants. For any weight parameter w _s , the nth ant moves from one point to the next according to the following probability formula (4):

In formula (4), i represents the i-th division scale value of the weight parameter w _s , and the value range of i is [1,101], Indicates the sum of the pheromones of all ants in the weight parameter w _s , j represents the jth ant, and the ant will update its own pheromone value according to the error in each iteration of the neural network, and the pheromone value in the present invention is One of the above 101 division scale values.;

The nth ant passes through the division scale value of the weight parameter w _s only once, and records the division scale value of the corresponding point. The combination of these points constitutes a set of weight parameters of the neural network;

(3c) Use the second-hand car transaction data and its vehicle data as input training samples, use the weight combination obtained in step (3b) as the parameters of the neural network, and use The Sigmoid S-type activation function performs the output calculation of the neural network, as shown in formula (5):

After obtaining the output of the neural network, calculate the mean square error, and take the maximum value of the mean square error, as shown in formula (6):

In formula (6), SampleNum is the number of samples, y and o are respectively the expected output value and the actual output value of the neural network;

(3d) After all ants construct the solution, record a set of weights with the smallest E, compare the minimum error E _min with ε _ACO , if E _min <ε _ACO , then directly complete the initialization process and exit, otherwise go to step (3e);

(3e) Pheromone update: The pheromone update strategy of the i-th point (division scale value) of the weight w _s is shown in formula (7):

In the formula, The i-th point of weight w _s corresponds to the updated pheromone value of the n-th ant in the t-th generation ant colony, and the range of μ is [10,100];

(3f) Steps (3b)-(3d) are repeated until the maximum number of iteration countmax is satisfied, and the initialization process is completed.

Preferably, the process of further training the optimized BP neural network in the step (4) is: a group of E _min minimum weight parameters found by the ant colony algorithm are used as the initial weight and deviation of the BP algorithm, and the network output is calculated and the actual output, and backpropagate the error from the output layer to the input layer, further adjust the weights and biases, and repeat the above process until the training exit condition is met.

In the traditional ACO pheromone update strategy, the offspring ants select the path probabilistically according to the pheromone table updated by the previous generation ants. Since the pheromone volatilization coefficient ρ is a fixed value, those paths that have never been searched The pheromones will gradually disappear, resulting in a decrease in the probability of these paths being selected, while the probability of non-optimal paths with increased pheromones will increase, and then the algorithm will fall into a local optimum. Therefore, in order to improve the global search ability of the algorithm and make the algorithm find the optimal solution as much as possible, it is necessary to appropriately reduce the number of pheromones on the selected path, that is, to appropriately weaken the pheromone contribution ability of offspring ants, and introduce an adjustment factor The value of μ is generally in the range of [10,100]. In this embodiment, μ takes 10.

In this embodiment, both the transfer function of the hidden layer and the activation function of the output layer in the BP neural network are Sigmoid S-type tangent curve activation functions.

The BP neural network training process involved in the present invention is a conventional technical means mastered by those skilled in the art, so it is not described in detail.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Both modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (5)

1. one kind is based on the improved BP neural network used car price evaluation algorithm of ant group algorithm, which is characterized in that choose three layers BP neural network is prototype, is optimized using weight initialization process of the improved ant group algorithm to BP neural network, is established Used car price evaluation model, the specific steps are as follows:

Step 1: data are acquired and are pre-processed；

Step 2: determining network topology structure；

Step 3: the weight initialization process of BP neural network being optimized using improved ant group algorithm；

Step 4: further the BP neural network after training optimization predicts used car price.

2. one kind according to claim 1 is based on the improved BP neural network used car price evaluation algorithm of ant group algorithm, It is characterized in that, the data of step (1) acquisition are second-hand vehicle data and its transaction data as sample data, including vehicle Number, productive year sell a year month, city, discharge capacity, new car price, mileage, vehicular applications, the degree of wear and individual and refer to Lead valence；The sample data of acquisition is pre-processed in the step (1): the sample data of acquisition is normalized, It specifically normalizes in [0,1] section, normalization formula is as follows:

X₀=(X-X_min)/(X_max-X_min) (1)

In formula (1), X₀For the data after normalization, X is initial data, X_max、X_minThe respectively maximum value of raw data set and most Small value.

3. according to claim 1 or 2 a kind of based on the improved BP neural network used car price evaluation calculation of ant group algorithm Method, which is characterized in that in the step (2) network topology structure be three layers of BP neural network structure, including input layer, hide Layer, output layer, wherein

Input layer: inputting pretreated sample data, and each input node is mapped to a vehicle association attributes, specifically For car number, productive year, sell a year month, city, discharge capacity, new car price, mileage, vehicular applications, the degree of wear and a People's guiding price totally 10 neurons；

Hidden layer: rule of thumb formula (2) estimates best hidden layer neuron number,

In formula (2), H is hidden layer neuron number, and I is input layer number, and O is output layer neuron number, and a is [1,10] constant in；

Output layer: output used car activity price is as a result, Real-time Feedback is for reference.

4. one kind according to claim 3 is based on the improved BP neural network used car price evaluation algorithm of ant group algorithm, It is characterized in that, the optimization process of the weight initialization in the step (3) is as follows:

The weight section [- 5,5] of BP neural network is evenly dividing as 100 equal parts, joins for each weight by (3a) parameter setting Number establishes a pheromones table, remembers w_sThe value range of the weighting parameter for needing to optimize for s-th, s is [1, N], and N is expressed as weighing The total number of value, i are to divide scale value, and every two divides adjacent scale value and constitutes a sub-regions；τ (i) is letter corresponding to i Breath element value, wherein weight section [- 5,5] are evenly dividing as 100 equal parts, then divide scale value and share 101；

The total number N of the weight is by input layer number I, hidden layer neuron number H and output layer neuron number O It determines, calculation formula is as follows:

N=H* (I+O+1)+O (3)；

Meanwhile setting information element initial value is τ (i)=C, C ≠ 0, pheromones volatility coefficient ρ, pheromones incremental intensity Q, ACO are most Big the number of iterations countmax and ACO optimize termination condition ε_ACO；

(3b) discharges m ant, for any weighting parameter w_s, n-th ant according to following new probability formula (4) from some movement To next point:

In formula (4), i indicates weighting parameter w_sI-th of division scale value, the value range of i is [1,101], Indicate weighting parameter w_sIn all ants the sum of pheromones；

N-th ant is from weighting parameter w_sDivision scale value by and merely through primary, record the division scale value of respective point, The combination of these points constitutes neural network weight parameter w_sOne group of weighting parameter；

(3c) using Second-hand Vehicle Transaction data and its vehicle data as input training sample, the weight group obtained using step (3b) Cooperation is the parameter of neural network, and the input layer of BP neural network is all made of Sigmoid S to hidden layer, hidden layer to output layer The output that type excitation function carries out neural network calculates, as shown in formula (5):

After SH obtains the output of neural network, mean square error is calculated, and take the maximum value of mean square error, as shown in formula (6):

In formula (6), SampleNum is number of samples, and y is desired output, and o is the real output value of neural network；

The smallest one group of weight of E is recorded after (3d) all ant construction solutions, compares minimal error E_minWith ε_ACOSize, if E_min<ε_ACO, then it is done directly initialization procedure and exits, otherwise goes to step (3e)；

(3e) Pheromone update: weight w_sI-th division scale value pheromone update strategy such as formula (7) shown in:

In formula (7),For weight w_sDivide the corresponding t of scale value i-th and passed through for n-th ant in ant colony Later the pheromones value updated, the value range of μ are [10,100]；

(3f) repeats step (3b)-(3d), until meeting maximum number of iterations countmax, completes initialization procedure.

5. it is according to claim 4 a kind of based on the BP neural network used car price evaluation algorithm for improving ant colony, it is special Sign is, the process of the BP neural network after optimization is further trained in the step (4) are as follows: according to ant group algorithm by step 3 (d) E found_minThe initial weight and deviation of the smallest one group of weight and deviation as BP algorithm calculate network output and reality Error between output, and error is propagated backward to input layer by output layer further adjusts weight and deviation, repeat more than Process, until meeting training exit criteria.