Evaluation of Innovation and Entrepreneurship Ability of Computer Majors based on Neural Network Optimized by Particle Swarm Optimization

— The current evaluation index systems (EISs) of innovation and entrepreneurship (I&E) ability are not sufficiently systematic, scientific, or practical. To solve the problem, this paper tries to evaluate the I&E ability of computer majors, using neural networks improved by particle swarm optimization (PSO). Firstly, an EIS of 22 second-level indexes under 5 first-level indexes was designed to evaluate the I&E ability of college computer majors. Next, an evaluation model was developed based on fuzzy neural network (FNN), and the corresponding training algorithm was created. Moreover, an improved PSO was introduced to optimize the FNN, and the optimization process was detailed. The proposed model was proved effective through experiments.


Introduction
The employment situation is quite severe in current society, and college graduates are under great employment pressure [1][2][3][4], so in order to help them find suitable jobs and provide them with more employment options, colleges and universities have set various innovation and entrepreneurship training programs and education courses, in the hops of cultivating innovative and entrepreneurial talents for the country [5][6][7][8].
Studies on I&E ability have accumulated fruitful results in recent years [9][10][11][12][13][14][15], for example, Su and Yu [16] explored the connotation and significance of the "Internet + education" model, and developed a new school-enterprise cooperation model. Louis Louw et al. [17] empirical tested the spatial spillover effects of regional economic development level on the I&E ability of the region, they constructed a Spatial Dubin Model (SDM) and verified the significant positive spatial correlation between regional I&E ability and regional economy. Chen [18] discussed the features and elements of college students' I&E ability, screened the indexes, and established an EIS for it, also, the weight values of each index were given in the paper. Kim et al. [19] introduced the experiencing-type deliberate practice training into college students' entre-preneurship and career guidance courses, combining with the experiments on the I&E ability of college students majoring in computer science, they proved that there're obvious differences in the students' I&E ability before and after implementing the I&E teaching mode based on creative problem-solving. Xu et al. [20] explored deep into the aspects of the integration of computer technology, the innovation mode, and the I&E ability training of computer professionals. For the purpose of cultivating students' I&E ability, Ran and Wang [21] optimized the teaching mode of computer network experimental courses. Scholar Li [22] proposed an all-around mechanism for cultivating innovative talents majoring in computer science in local colleges and universities, and integrated it into the construction of professional curriculum and the talent training mode.
Existing EISs of I&E ability are not sufficiently systematic, scientific, or practical [23][24][25][26][27], therefore, this paper attempts to research the evaluation of I&E ability of college computer majors and propose a new idea to optimize the neural network model using PSO. The second chapter of this paper constructs an EIS of I&E ability of college computer majors containing five first-level indexes of innovation ability, professional quality, entrepreneurial resource integration, team organization and management, and entrepreneurial risk-taking. The third chapter constructs an FNN-based evaluation model and gives the corresponding training algorithm. The fourth chapter elaborates on the optimization method of the FNN optimized by an improved PSO; the last chapter uses experimental results to verify the effectiveness of the proposed model.

2
The proposed EIS After fully and comprehensively considering the differences between computer majors and students of other majors, this paper designed an EIS of I&E ability of computer majors based on a few principles such as the comprehensive principle, the scientific principle, and the operable principle, etc. The professional ability of computer majors can be described from aspects such as whether a student has obtained certificates in software-related exams, and whether the student has software patents or copyrights, etc. The practical ability of computer majors can be measured by the index of whether a student has won awards in computer skill competitions. This paper mapped the student's personal computer education experience to their career planning direction and I&E ability, and used 5 first-level indexes (innovation ability, professional quality, entrepreneurial resource integration, team organization and management, and entrepreneurial risk-taking) to explain the main evaluation content of the I&E ability of college students majoring in computer science, the details are given in Figure 1.
As can be seen from the figure, the first-level index innovation ability IE1 contains 3 second-level indexes: innovation awareness IE11, learning ability of professional knowledge of computer technology IE12, practical operation ability of computer technology IE13; the first-level index professional quality IE2 contains 8 second-level indexes: braveness IE21, self-confidence IE22, toughness IE23, integrity IE24, positivity IE25, optimism IE26, execution ability IE27, sense of responsibility IE28; the first-level index entrepreneurial resource integration IE3 contains 3 second-level indexes: entrepreneurial resource requirement analysis ability IE31, entrepreneurial resource allocation ability IE32, entrepreneurial fund raising ability IE33; the first-level index team organization and management IE4 contains 4 second-level indexes: interpersonal communication ability IE41, leadership ability IE42, teamwork ability IE43, team operation ability IE44; the first-level index entrepreneurial risk-taking IE5 contains 4 second-level indexes: computer industry entrepreneurial risk recognition ability IE51, computer industry entrepreneurial risk resistance ability IE52, stress resistance ability IE53, self-regulation ability IE54.

FNN-based evaluation model and training algorithm
To construct a FNN for evaluating the I&E ability of computer majors, the first step was to build a neural network with multiple hidden layers for storing the fuzzy control language and sentences, then, through offline training of the network, the connection weight values were updated to realize the fuzzy reasoning of the I&E ability of computer majors. Figure 2 shows the structure of the constructed FNN. Assuming the constructed 5-layer network has m inputs and e outputs, then the node number of each layer could be determined according to the following method: For the input layer nodes of the network, their number was set to be the same as the number of evaluation index samples, respectively were al, a2, ..., am, and there're m nodes in total.

Fig. 2. Structure the constructed FNN
Assuming: every input evaluation index sample has n fuzzy language variables, then, the number of neuron nodes in the fuzzy language layer of the network is n×m. Since the input layer is directly connected to the fuzzy language layer, the membership degree of the fuzzy language variables can be judged according to the input value of each evaluation index sample based on the Gaussian function. Assuming: λi j represents the membership degree function of the i-th neuron node of the input layer on the j-th fuzzy language variable, and HEij and εij represent the central value and width of the membership degree function, then, there is: For the fuzzy rule layer that is used to describe the fuzzy rules, usually, its node number is set to be the m power of n, namely n m . Formula 2 gives the calculation formula of the fitness degree βj of the fuzzy rules: As for the normalization layer that is used to normalize the fitness degree of the fuzzy rules, its neuron node number was set to be the same as the number of nodes in the fuzzy rule layer, namely n m . Formula 3 gives the formula for the normalization processing of this layer: (3) http://www.i-jet.org

Paper-Evaluation of Innovation and Entrepreneurship Ability of Computer Majors based on Computer…
The final output of the network output layer that realizes clear operations is the evaluation result of the I&E ability of computer majors. Assuming: θ represents the connection weight of the network, then Formula 4 gives the formula for the clear operations of this layer: The three variables can be expressed as a vector shown as Formula 5: Since the numbers of input and output neuron nodes of the constructed FNN model were fixed, and the fuzzy language number was also a fixed value n, then, in the model, the parameters that need to be updated via training included the connection weight θ, and the central value HEij and width εij of the membership degree function. The network structure of the constructed FNN model was a feedforward-type network, so it could simulate the forward propagation process of the error of BP neural network to perform iterative operations on θ, HEij and εij.
First, based on the forward propagation of the input sample data, the input and output of each layer of the neural network were calculated; assuming: gi (r) and hi (r) represent the input processing and activation processing of the i-th neuron in the r-th layer of the neural network, and ai (r) represents the output of the i-th neuron in the r-th layer of the neural network, then the output and input of the input layer satisfied the following equations: Assuming: for the i-th neuron, there're ni corresponding fuzzy language, the value range of j is [1, ni], then the output and input of the fuzzy language layer satisfy: Since the number of fuzzy rule sentences was n m and the value range of j was [1, n m ], the output and input of the fuzzy rule layer satisfied the following equations: The value range of j of the normalization layer was also [1, n m ], and the output and input of this layer satisfied the following equations: The value range of output layer i was [1, e], the output and input of this layer satisfied: After the evaluation of the I&E ability of computer majors was output by the output layer, it could be compared with the expected evaluation result. The adjustment of all neural network parameters such as θ, HEij and εij was to ensure the smallest output error. Assuming: φi represents the expected value of the i-th neuron in the output layer, bi represents the actual value of the i-th neuron in the output layer, M represents the size of the training data, then, Formula 11 gives the function for quantifying the errors of each output neuron: The neural network training method used in this paper referred to the error forward propagation of the BP neural network, this method adjusted parameters θ, HEij and εij based on the correction of the calculation results of propagation error of each layer. Formula 12 gives the calculation formula of the output layer error: The connection weight θ between the normalization layer and output layer neurons needs to be adjusted. Based on the propagation error of the output layer, the increment of θ can be calculated using Formula 13: Formula 14 gives the calculation formula of the error that is forwarded to the normalization layer: Formula 15 gives calculation formula of the error that is forwarded to the fuzzy rule layer: Formula 16 gives the calculation formula of the error that is forwarded to the fuzzy language layer: Considering that the central value HEij and the width εij of the membership degree function were both in the fuzzy language layer, Formula 17 gives the calculation formula of the increment of HEij when the error is forwarded to the fuzzy language layer: Formula 18 gives the calculation formula of the increment of width εij of the corresponding membership degree function: Formula 19 shows the formula for adjusting the connection weight θ during the training process of the neural network: Formula 20 gives the formula for adjusting the central value HEij of the membership degree function: Formula 21 gives the formula for adjusting the width εij of the membership degree function:

The FNN optimization method based on improved PSO
This paper adopted an improved PSO to optimize the constructed FNN. The improved PSO realizes the balancing between global optimal value searching and local optimal value searching of the algorithm and ensures accurate optimal value searching and fast convergence by adjusting the speed of the particle swarm and the learning factors.
To improve the algorithm' ability to search for global and local optimal values, this paper introduced the inertia weight χ that affects the speed adjustment into the speed adjustment process, Formula 22 gives the formula for adjusting the speed of the particle swarm: In fact, χ characterizes the size of the impact of particle speed at moment φ on the speed of the particle at moment φ+1. The inertia weight χ in this paper adopted an adaptive linear adjustment strategy. The PSO needs to increase the search space of the particle swarm in the initial execution stage to find the location of optimal solution of the particle swarm faster, at this time, it needs to set a larger χ value. In the later execution stage, the PSO needs to enhance the particle swarm's ability to search for local optimal solution, and then find the accurate location of the optimal solution of the particle swarm faster, at this time, it needs to set a smaller χ value. Assuming: χC represents the inertia weight in the initial execution stage of the algorithm and its value range is [0.9,1.2]; χF represents the inertia weight in the later execution stage of the algorithm and its value range is [0.4,0.8]; ψ represents the number of iteration times; φ represents the number of current iterations; Formula 23 gives the formula for adjusting the inertia weight χ: In the learning factors, the cognition factor describes the degree of cognition of each particle to its own position, and the social factor describes the impact of the overall position of the particle swarm to each particle. Wherein, fitting has a great impact on the individual particles, and the latter is conducive to conducting local search in the region where the optimal solution is located. In order to effectively improve the accuracy and convergence speed of the algorithm, this paper adjusted the cognition factor HE1 and the social factor HE2 in Formula 29. Assuming HE1C and HE1F respectively represent the initial value and final value of the cognition factor, Formula 24 gives the adjustment formula for HE1: According to above formula, HE1 showed a decreasing trend; assuming HE2C and HE2F respectively represent the initial value and final value of the social factor, Formula 25 gives the adjustment formula for HE2: According to above formula, HE2 showed an increasing trend. Parameters of the neural network such as θ, HEij, and εij that need to be optimized were coded in matrix form, then, each particle in the improved PSO can be expressed as: Figure 3 shows the implementation flow of the FNN optimized by the improved PSO. http://www.i-jet.org

Paper-Evaluation of Innovation and Entrepreneurship Ability of Computer Majors based on Computer…
Start Determine the structure of FNN Initialize the parameters of the neural network and the particle swarm model Adjust the network parameters based on the optimal state of the particle swarm Compare the particle fitness before and after parameter adjustment, and update the optimal individual and optimal population Is the number of iteration times greater than the number of maximum iteration times Output the optimal parameters of the FNN Calculate the initial fitness of the particles and determine the optimal individual and optimal population Linear adjustment of inertia weight χ Adjustment of learning factors HE1 and HE2 Calculate the fitness of the particle and update the optimal individual and optimal population End Is the number of adjustment times greater than number of adjustment times of the optimal location

Experimental results and analysis
To ensure certain reliability and validity of the constructed EIS, this paper used a commonly-recognized reliability analysis method to analyze the data of the EIS. Table 1 gives the reliability test results of the constructed EIS. According to the table, the Cronbach's α values of the selected evaluation indexes of the I&E ability of computer majors were all greater than 0.80, and the total correlation coefficients of the corrected terms were also greater than 0.5.
In the simulation experiment designed in this study, a 5-dimensional FNN model structure was selected, that is, the error and error change rate between the predicted values and the actual values of the evaluation scores of the five first-level indexes (namely innovation ability, professional quality, entrepreneurial resource integration, team organization and management, and entrepreneurial risk-taking) were selected, then, by converting the quantified input factors of the two to the fuzzy theory domain, 5 fuzzy language variables were obtained, Figure 4 gives a schematic diagram of the selected membership function.
To verify the optimization effect of the improved PSO on the FNN, this study designed an experiment to compare the traditional PSO and the improved PSO. Suppose: the population size of the particle swarm was 100, the inertia weight χ was set to 0.9, and the initial values of the cognition factor HE1 and the social factor HE2 were set to 2.5 and 0.5, respectively. In the improved PSO, with the increase of the particle update times, parameters χ, HE1 and HE2 were adjusted gradually, the value of χ was decreased to 0.4, the value of HE1 was decreased to 0.5, and the value of HE2 was increased to 2.5. The maximum numbers of adjustment times and iteration times of the PSO were respectively set to be 10 and 200, and the maximum number of iteration times of the neural network was set to be 1000. Figure 5 and Figure 6 give the change curves of the training performance index of the FNNs improved by the traditional PSO and the improved PSO.   By comparing Figure 5 and Figure 6, we can see that the value of the training performance index of the FNN optimized by the improved PSO changed from about 0.03 before optimization to about 0.01 after optimization, and the overall convergence trend of the performance index curve of the network optimized by the improved PSO was better. Therefore, in the subsequent experiment on the evaluation effect of the I&E ability of computer majors, the parameters after the optimization of the improved PSO were set to be the initial values of the neural network. Figure 7 shows the fitness curve of parameter optimization of the improved PSO in the model evaluation process.

Conclusion
This paper researched the I&E ability of computer majors based on neural network optimized by PSO. In the paper, an EIS of 22 second-level indexes under 5 first-level indexes was designed to evaluate the I&E ability of college computer majors, and a corresponding FNN evaluation model was established. Then, the optimization method of FNN based on the improved PSO was elaborated, the reliability test results of the EIS were given through experiment, and the reliability and effectiveness of the constructed EIS were verified. Moreover, the paper also gave the change curves of the training performance index of the FNNs optimized by the traditional PSO and the improved PSO, and the results suggested that the overall convergence trend of the performance index curve of the network optimized by the improved PSO was better. At last, this paper compared the predicted results of the training set of the evaluation model and the subjective evaluation values given by the experts, which had further verified the effectiveness of the model in evaluating the I&E ability of computer majors.