Emotion recognition is a research hotspot in the field of artificial intelligence. If the human-computer interaction system can sense human emotion and express emotion, it will make the interaction between the robot and human more natural. In this paper, a multimodal emotion recognition model based on many-objective optimization algorithm is proposed for the first time. The model integrates voice information and facial information and can simultaneously optimize the accuracy and uniformity of recognition. This paper compares the emotion recognition algorithm based on many-objective algorithm optimization with the single-modal emotion recognition model proposed in this paper and the ISMS_ALA model proposed by recent related research. The experimental results show that compared with the single-mode emotion recognition, the proposed model has a great improvement in each evaluation index. At the same time, the accuracy of emotion recognition is 2.88% higher than that of the ISMS_ALA model. The experimental results show that the many-objective optimization algorithm can effectively improve the performance of the multimodal emotion recognition model.

It is worth mentioning that rainfall is the only input element source for the hydrologic cycle. However, excessive rainfall and the scarcity of it on Earth affect the tremendous flooding and severe drought that occur over short and long intervals [

Weather forecast can be divided into three types: short-term weather forecast [

Based on the above analysis, a new weather prediction model based on the improved quantum genetic algorithm (IQGA) [

As far as the current social development is concerned, we are all in the background of big data, and we cannot live without the influence of the data system. In the fierce competition process of all walks of life, the results of data mining often affect the decision-making direction of the whole project and then affect the final implementation and development results. It can be seen that data mining has become the focus of attention, and the data preprocessing which plays an important role in the process of data mining has also been further studied and found. Data preprocessing [

The complex high-dimensional dataset used in this study includes the daily weather forecast of many Australian meteorological stations in 10 years. There are about 140,000 pieces of data points, each of which has 22 characteristics, 21 columns are characteristic matrices (including time, place, maximum temperature of the day, rainfall, and wind speed), and the last column is tag (indicating whether it will rain tomorrow). In order to forecast the weather, we need to build the model forecast tag in the 21-column feature. The feature matrix of the dataset is composed of a part of classified variables and a part of continuous variables. For example, although the degree of cloud cover is expressed by numbers, it is a classified variable in essence. Most of the features are collected natural data, such as evaporation, sunshine time, and humidity, while a few of the features are composed of humans. There are also variables that simply represent sample information, such as the location and time of collection.

In the data preprocessing of this study, we will divide the training set and the test set first and then start the data preprocessing. This is because the test set is often impossible to get in advance in reality, and we do not want any process of our modeling to be affected by the test set data; otherwise, it is equivalent to telling part of the predicted answers of the model in advance. In this study, all operations we performed on the training set must be performed on the test set once.

After obtaining the complex real weather dataset, we first observe the composition of the dataset as shown in Table

Partial weather dataset.

Date | Location | MinTemp | MaxTemp | Rainfall | Sunshine u̇u̇ | Humidity | RainTomorrow |
---|---|---|---|---|---|---|---|

2014/8/26 | Launceston | 3.5 | 18.4 | 0 | 2.5 | 66 | No |

2013/5/1 | Perth | 14.4 | 25.7 | 0 | Null | 38 | Yes |

2017/5/9 | Wagga | 1.4 | 19 | 0 | 0.6 | 32 | No |

2016/6/11 | Uluru | 7.2 | 17.4 | 0 | 0.8 | 40 | No |

2010/3/26 | Cairns | 23.2 | 28.3 | 4.4 | 1.3 | 93 | Yes |

By observing the aforementioned dataset, we can find that there are a lot of empty data in the dataset, and there is also a case of dimensional inconsistency. At the same time, the data are composed of classification variables and continuous variables which need our data processing.

In this study, we first process the characteristics of the data. By observing the data, we can know that all data include 21 characteristics: time, place, temperature, humidity, cloud cover, rainfall, etc. They also include our tag: whether it will rain tomorrow. Through data visualization [

Year rainfall trend.

Monthly rainfall trend.

Location rainfall trend.

Through data visualization, we can intuitively find that the year has little impact on the tag, but it will increase the complexity of the data, but the month will have a great impact on the tag, and the location has a very obvious impact on the tag. In this way, these data can help the model to make a better and more accurate judgment on the label. So, we cannot delete it directly; we need to deal with it. The following ingenious processing of weather data is made in this study.

The month in time is extracted; because through data visualization analysis and common sense, we can know that the month will affect the probability of rain.

Does the day in time affect the tag? On the surface, it may not have any effect, but if it rains today, will the probability of rain tomorrow increase? Thinking of this, We have added new features to the dataset about whether it is raining today. We can set the threshold value by using today’s precipitation. In this study, we set rainfall greater than 1 mm as today’s rain and set the rest as no rain. After treatment, we add a new feature of whether it rains today.

We know that the climate corresponding to the location will affect our tag, so this study cleverly corresponds the location to the local climate, and we found the Australian land as shown in Figure

City climate map of Australia.

This is produced by the Australian Meteorological Agency and the Australian Building Code Board (ABCB), which divides the climate regions of different cities in different regions of Australia. The distribution map is divided into eight regions, which is very suitable for us to classify. If we can transform 49 locations into eight different climates, this information should be useful for judging whether it rains or not. Based on the data of the Meteorological Bureau and ABCB, we have made the corresponding climate-type data of the main cities in Australia and saved them in the csv-1 file, but we found that the location in the original weather dataset does not represent the city but the name of the weather station. We need to convert the name of the weather station in the dataset into the name of the city. We use the shortest direct distance between two points to convert. Next, we do a web crawler [

After finishing the above processing, we will fill in the missing value, code the type variable, and make the continuous variable dimensionless. Since our eigenmatrix is composed of two types of data, type and continuous, we must adopt different strategies to fill the missing values. In this study, if the feature is classified, we use mode to fill. If the feature is continuous, we use mean to fill. At this time, because we have divided the training set and the test set, because our test set may not be many pieces of data in reality, may be our test set has only one piece of data, and a certain feature is null, at this time, the mode of the test set itself does not exist, so in order to avoid this situation, we usually assume that the data distribution and properties of the test set and training set are similar, so we use the mode and mean of the training set to fill the test set. The normalization method is used in the dimensionless treatment of continuous variables.

After completing the above operations, our data preprocessing work is finished completely, and then we start to model.

Quantum genetic algorithm (QGA) [

The smallest information unit in QGA is represented by qubits. A qubit can not only represent two states of 0 and 1 but also represent any superposition state between them. Therefore, the state of a qubit can be expressed as

Therefore, a qubit can store and express information of two states at the same time. A qubit string with a length of

The evolution of quantum genetic algorithm is to use the quantum gate to rotate the quantum state, which is another significant difference from classical genetic algorithm. The transformation of the quantum logic gate is unitary transformation in the Hilbert space, which can ensure that the superposition state after change still meets the normalization condition.

In the quantum genetic algorithm, the quantum rotating gate [

The rules for

Among them,

Figure

Diagram of rotation variation.

The quantum rotation gate completes the update operation by changing the amplitude of the quantum superposition state. The change of the amplitude changes the probability of each basic state obtained after measurement while maintaining the diversity of the population. The initial population will gradually deflect to the species with higher adaptability after the quantum rotation gate operation and finally converge to the optimal population to find the optimal solution of the problem.

In this paper, the traditional quantum genetic algorithm is improved in three aspects, and an improved quantum genetic algorithm is proposed. Next, three improvements are introduced.

(1) Improvement of the quantum population initialization adjustment strategy: quantum genetic algorithm uses the quantum superposition state to represent individuals in the population. Therefore, compared with the classical genetic algorithm, it maintains the same population diversity and greatly reduces the population size. In general, the population initialization of quantum genetic algorithm uses formulas to encode individuals, and each gene collapses to

In order to make the initial population more evenly spread over the whole solution space, the initial population can be initialized by grouping. When the traditional quantum genetic algorithm uses the formula to encode the quantum state, the parameter

Therefore, in this study, the initial population is divided into five subpopulations, and the population initialization rules are as follows:

The above formula divides the initial population into five parts so that the initial population can spread over the whole solution space.

This initialization method can make the initial population spread over the whole solution space and avoid the disadvantage that the probability amplitude of the quantum basic state is too average which may result in the measurement results concentrated in a certain interval. It can search from multiple local solution spaces at the same time, ensure the diversity of chromosomes, and improve the convergence rate of the algorithm.

(2) Improvement of the rotation angle adjustment strategy: the core of quantum genetic algorithm is the operation of update evolution which uses quantum rotation gate to rotate the quantum state at a certain angle. Premature convergence may occur when the rotation angle is too small; the convergence rate is too slow when the rotation angle is too large. Therefore, how to choose the rotation angle of the quantum rotary gate will directly affect whether the quantum genetic algorithm can find the optimal solution quickly and effectively. However, the rotation angle of the quantum rotation gate selected by the update strategy is usually fixed in the standard quantum genetic algorithm, or the evolution operation is completed according to the update strategy table. The two rotation angles are relatively fixed, which makes the algorithm not flexible enough to play the best performance of the quantum genetic algorithm.

Our improved quantum genetic algorithm uses a dynamic adjustment method for the rotation angle of the quantum rotation gate. When the current number of iterations is small, it indicates that there is still a large optimization space for the current quantum individual as a whole, and large angle rotation is required to achieve faster convergence. When the current number of iterations is closer to the maximum number of iterations, it shows that the whole quantum individual has become more and more excellent, so we only need to gradually reduce the rotation angle to find the best individual.

The dynamic adjustment strategy for the rotation angle

Among them,

New quantum rotation gate strategy.

0 | 0 | False | −1 | +1 | 0 | ||

0 | 0 | True | −1 | +1 | 0 | ||

0 | 1 | False | +1 | −1 | 0 | +1 | |

0 | 1 | True | −1 | +1 | 0 | ||

1 | 0 | False | −1 | +1 | 0 | ||

1 | 0 | True | +1 | −1 | 0 | ||

1 | 1 | False | +1 | −1 | 0 | ||

1 | 1 | True | +1 | −1 | 0 |

(3) Selection strategy improvement: the new mechanism of probability selection is adopted to abandon the traditional strategy of keeping excellent individuals and discarding individuals with low adaptability, and the probability of being selected can be determined flexibly according to the size of individual adaptability, which not only ensures the population richness but also effectively optimizes the algorithm into the local optimal solution, and each selection makes the algorithm move towards a good evolutionary direction. The specific implementation form is as follows:

Support vector machine [

SVM is developed from the optimal classification surface in the case of linear separability. The basic idea of the SVM is to map the input vector to a high-dimensional feature space through a preselected nonlinear mapping and turn the required nonlinear problem into a linear separable problem in the high-dimensional space. In this high-dimensional space, an optimal linear classification surface is constructed which can separate the two types. Moreover, the classification interval between the two categories is maximized.

Set the sample set to

Then,

The Lagrange function is constructed:

The partial differentiation of

After solving

For the nonlinear classification, the kernel function

Then, the corresponding classification function is

Considering that some samples may not be classified correctly, in order to ensure the correctness of classification, the relaxation factor

The optimization objective function becomes

At present, there are four kinds of kernel functions that are widely used:

Linear kernel:

Poly kernel:

RBF kernel:

Sigmoid kernel:

An important step is to select kernel function and kernel parameter when the SVM is used in weather forecast. Vapnik et al. found that different kernel functions have little effect on the SVM performance, but the parameters of kernel function and error penalty factor

Support vector machines have good generalization ability. They can solve the problems of small sample and nonlinearity and overcome the shortcomings of traditional methods, such as dimension disaster and overlearning. However, there are many aspects that affect the recognition accuracy of the SVM including the selection of kernel function which is not only often related to human experience but also to sample data. It is impossible for human experience to be effective for all sample data. There are many optimization methods for parameter variables in practical problems, among which search algorithm is widely used. It searches for the whole parameter variable space. It is necessary to select an effective search algorithm to obtain the optimal parameters considering the large parameter space and the limited calculation time. Aiming at the weather prediction system studied in this paper, the parameter optimization of the traditional SVM classification model has reached the goal of improving the classification accuracy. Based on this, the IQGA-SVM classification model is constructed. The optimization ability of the improved quantum genetic algorithm is discussed in the following.

The quantum genetic algorithm starts from the initial population which is composed of the whole solution space with average distribution to search in the parameter space. The individual state is represented by the quantum bit probability amplitude, which enhances the diversity of the population and expands the optimization ability. Individuals with higher adaptability are selected for population renewal through the selection operation in genetic operation. Population renewal is carried out through the quantum rotation gate. Population renewal means that there are always new individual transformations which not only enhance the population diversity but also expand the scope of search and can obtain the global optimal solution better than genetic algorithm. Quantum genetic algorithm only needs to calculate the fitness value. Therefore, it is convenient to connect the SVM as an intermediate variable for parameter optimization.

IQGA-SVM algorithm flow is as follows:

First, a predesigned angle sequence is generated by initialization. Then, it is transformed into a population of

Construct

Use the fitness function to calculate each individual in

Judge whether the specified precision termination condition is triggered. If not, continue the following operations. If the condition is met, output the optimal parameter and construct the optimal parameter model.

According to the number of iterations, the rotation angle of the quantum rotation gate is changed dynamically.

Determine whether the set upper limit of iteration times is reached. If the upper limit is reached, output the optimal parameters, and construct the optimal parameter model. The algorithm flow of the weather prediction model based on IQGA-SVM is shown in Figure

IQGA-SVM algorithm flow.

In order to verify the feasibility of the model, the partial random sampling of the weather dataset is carried out, and the simulation experiment is carried out on software Jupyter Notebook, and Python 3.7 version is used for testing.

In order to reduce the search range of the improved quantum genetic algorithm, the general range of the parameters to be adjusted is determined by the traditional learning curve of parameter adjustment. We found that the tag of the weather dataset has the problem of sample imbalance after data exploration. The proportion of rainy weather and nonrainy weather is about 3 : 1, so we need to add a

Learning curve of linear kernel function

Learning curve of linear kernel function class weight.

Through observation, it can be found that, for IQGA-SVM, the model of the weather forecast dataset in this study, the penalty coefficient

IQGA parameter and kernel function to be found parameter range.

IQGA parameter value | SVM value of kernel function | |||||
---|---|---|---|---|---|---|

Parameter | Value | Kernel function | ||||

Population number | 100 | Parameter | Linear | Poly | rbf | Sigmoid |

Chromosome number | 4 | No | ||||

Chromosome length | 20 | No | 1 | No | No | |

Cofe0 | No | No | ||||

iter_num | 30 | |||||

After determining the range of the parameters to be searched, the improved quantum genetic algorithm is used to optimize the parameters of the four kernel functions, respectively, and the optimization effect is shown in Figure

SVM kernel function comparison.

Through the analysis of Figure

Search result table.

Kernel | Cofe0 | ||||
---|---|---|---|---|---|

Linear | 0.4759025573730469 | No | No | No | 0.8845815658569336 |

Poly | 1.9044914245605469 | 0.0014705657958984375 | 8.277586 | 1 | 1.908930778503418 |

rbf | 2.34393310546875 | 0.00496673583984375 | No | No | 1.8635177612304688 |

Sigmoid | 0.7698040008544922 | 0.0029439926147460938 | 6.24142 | No | 1.2771968841552734 |

Next, IQGA-SVM model (poly kernel function), support vector machine model based on genetic algorithm (GA-SVM), and support vector machine model based on improved ant colony optimization (ICMPACO-SVM) [

Comparison of the algorithm effect.

Algorithm performance comparison.

Model | Average accuracy (%) | Optimal accuracy (%) | |
---|---|---|---|

Random forest | — | 85.4 | 0.85 |

Decision tree | — | 85.0 | 0.85 |

Logistic regression | — | 84.7 | 0.84 |

XGBoost | — | 86.0 | 0.86 |

SVM | — | 85.7 | 0.85 |

GA_SVM | 86.5 | 86.7 | 0.87 |

QGA_SVM | 86.7 | 87.0 | 0.86 |

ICMPACO_SVM | 86.8 | 87.1 | 0.88 |

IQGA_SVM | 87.1 | 87.3 | 0.88 |

ROC curve of the poly kernel function.

Next, we introduce the parameter setting of several other parameter searching algorithms:

Grid search: the parameters include

Random walk: the parameters include steps = 0.5, precision control (epsilon = 0.00001), number of variables (var_num = 4), initial point (

GA: parameter search scope is the same as grid search. Population size

ICMPACO: parameter search scope is the same as grid search. Ant number = 50, pheromone factor

The IQGA-SVM is compared with the commonly used parameter adjustment methods to verify its advantages in accuracy and efficiency. Grid search is the most commonly used method of parameter adjustment in the industry. Its idea is exhaustive search, traversing all the candidate parameters to find the best value of the parameters to be selected. Its disadvantage is that it takes a long time and can only be selected in the given candidate set. Random walk is a global optimization algorithm. According to the law of large numbers, as long as the random number is enough, the optimal or near-optimal parameters can always be found. Its disadvantage is that it highly depends on the initial value, and the random results may be inconsistent. The idea of quantum genetic algorithm is the survival of the fittest. Keep some excellent individuals in each iteration, and there is a certain probability of crossover and mutation, resulting in new samples. Therefore, quantum genetic algorithm is also a global optimization algorithm. When the mutation probability is large, quantum genetic algorithm will degenerate into random walk. IQGA-SVM, QGA-SVM, GA-SVM, grid search, and random walk were used to adjust parameters for 30 times, respectively. The comparison results are shown in Table

Comparison of parameter adjustment methods.

Method | Average accuracy (%) | Optimal accuracy (%) | AUC | Running time (s) | Parameter list |
---|---|---|---|---|---|

Grid search | 85.3 | 85.8 | 0.862 | 3056 | [2.23, 0.05, 3.3, 1, 1.3] |

Random walk | 85.5 | 86.2 | 0.866 | 1827 | [1.78, 0.007, 6.53, 1, 1.56] |

GA-SVM | 86.5 | 86.7 | 0.870 | 1235 | [2.19, 0.002, 7.23, 1, 1.91] |

QGA-SVM | 86.7 | 87.0 | 0.870 | 1302 | [1.97, 0.005, 8.84, 1, 1.51] |

ICMPACO-SVM | 86.8 | 87.1 | 0.880 | 2389 | [1.823, 0.0008, 8.5, 1, 1.45] |

IQGA-SVM | 87.1 | 87.3 | 0.880 | 1364 | [1.90, 0.001, 8.3, 1, 1.56] |

Table

The experimental results show that IQGA-SVM is better than linear regression and random forest algorithm in prediction accuracy. It is better than grid search and random walk in running time and parameter adjustment. Compared with the GA-SVM algorithm, IQGA-SVM can improve the prediction accuracy at a sacrifice of a small amount of time.

In this paper, the real complex weather dataset is selected as the experimental data through ingenious data preprocessing to enhance the potential of data mining. Then, the traditional quantum genetic algorithm is improved from several directions which effectively enhances the global search ability and efficiency of the quantum genetic algorithm. Next, the effects of IQGA-SVM, QGA-SVM, GA-SVM, and other classical machine learning algorithms on the dataset are compared. At the same time, IQGA-SVM is compared with grid search, random walk, QGA-SVM, and GA-SVM model in a variety of evaluation indicators. The advantages of the improved quantum genetic algorithm in parameter optimization are verified from several aspects. The experiment shows that the model of support vector machine based on the improved quantum genetic algorithm established in this paper has the optimal prediction effect. Through data analysis, the improved quantum genetic algorithm effectively improves the performance of the classifier and the accuracy of the weather prediction system by optimizing parameters. At the same time, we also find that the search range of IQGA-SVM will gradually decrease in the late iteration and occasionally fall into the local optimal solution. This leads to poor optimization effect. Our analysis shows that it may be because a large number of individuals are concentrated in the local optimal solution and cannot jump out at the end of the iteration. In the next study, we will consider adding the dynamic catastrophe strategy to solve this problem and further improve the optimization effect. This study has good academic research and engineering application value. The main contribution of this paper is to propose a new parameter optimization method. In the next research, we will further optimize the effect of quantum genetic algorithm. At the same time, it is combined with more machine learning algorithms to increase its scalability. It can improve the efficiency and accuracy of optimization by using the parallelism and good global search ability of quantum algorithm. It can also be applied to the weight and bias optimization of the neural network, so as to improve the learning ability of the neural network.

The data used in this article are available at

The authors declare that they have no conflicts of interest.

This research was supported by the National Natural Science Foundation of China (nos. 61772295, 61572270, and 61173056), the PHD Foundation of Chongqing Normal University (no. 19XLB003), the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant no. KJZD-M202000501), Chongqing Technology Innovation and Application Development Special General Project (cstc2020jscx-lyjsA0063), and Chongqing Normal University Graduate Scientific Research Innovation Project (Grant no. YKC20042).