A Novel Network Architecture of Decision-Making for Self-Driving Vehicles Based on Long Short-Term Memory and Grasshopper Optimization Algorithm

Long short-term memory network is one of the most important network architectures of decision-making for self-driving vehicles. Nevertheless, the decision-making accuracy of long short-term memory network is limited, the information of the surrounding vehicles is not taken into consideration, which is critical for the decision-making of the ego vehicle, and the classification capability of long short-term memory network is poor. In this article, a novel network architecture called improved long short-term memory network with support vector machine classifier optimized by grasshopper optimization algorithm (GOA-ImLSTM) is proposed. Three improvements are presented in GOA-ImLSTM. Firstly, to consider the information of the surrounding vehicles, a new network architecture, used to extract vital features for self-driving vehicles, with three parallel long short-term memory network units and a network unit serial connected according to vehicle location is designed. Secondly, to improve classification accuracy, support vector machine with stronger classification capability than softmax is introduced to accomplish the classification task. Thirdly, to promote the classification capability of support vector machine, grasshopper optimization algorithm is employed to optimize the parameters of support vector machine. Moreover, to balance exploration and exploitation ability of grasshopper optimization algorithm, dynamic weights in position movement formula are defined. The experiments indicate that GOA-ImLSTM improves the accuracy of results compared with other decision-making methods for self-driving vehicles on the Next Generation SIMulation.


I. INTRODUCTION
With the continuous breakthrough of artificial intelligence, autonomous vehicles have moved towards practicality [1]. The self-driving system is a highly autonomous system, including perception module [2], path planning module [3], behavior decision module [4], and adaptive control module [5]. The behavior decision module is the key technology The associate editor coordinating the review of this manuscript and approving it for publication was Vivek Kumar Sehgal .
to determine the safety and stability of autonomous vehicles [6]. Recently, research on driving decision-making for autonomous vehicles has made great strides. Main decisionmaking algorithms are divided into three categories, rulebased methods, reinforcement learning methods [7], and deep learning methods [8].
For rule-based methods, making autonomous driving decisions by self-definition rules on modeling or the features extracted by neural networks as rules. Furda and Vlacic [9] used multiple criteria decision making (MCDM) to select the VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ most appropriate driving maneuver from a set of feasible driving maneuvers, which are obtained through Petri net on vehicle environment model. Chong et al. [10] proposed a fuzzy rule-based neural network model to obtain driver behavior rules from individual vehicle trajectory data. Barman et al. [11] designed a fuzzy inference system to generate fuzzy rules and define a nonlinear mapping of input data. Li et al. [12] chose T-S fuzzy neural network (TSFNN) to establish driving decision-making model under emergency situations. Despite the high accuracy of simple scenarios in the rulebased methods, the complexity of the rule-making method limits its further development in complicated environments. For reinforcement learning methods, obtaining optimal strategies by maximizing long-term future rewards. Ngai and Yung [13] employed a multiple-goal reinforcement learning framework to determine action decisions by considering seven different goals. Gindele et al. [14] proposed a novel approach, which is based on a continuous partially observable Markov decision process (POMDP), to obtain decision-making for autonomous driving. Song et al. [15] first developed a continuous hidden Markov model to predict the motion intention, then built a POMDP to model the general decision-making framework. You et al. [16] modeled the autonomous vehicle as a stochastic Markov decision process (MDP) and considered the driving style of an expert driver as the target to be learned by reinforcement learning and inverse reinforcement. Chen et al. [17] proposed a fuzzy Markov prediction model which can estimate the short-term traffic conditions in urban environment. Although the reinforcement learning methods made a safe and efficient driving decision in uncomplicated condition, the complexity of truly driving environments makes the accuracy of decision making is limited.
For deep learning methods, many deep networks are used in driving decision-making, such as convolutional neural network (CNN) [18], recurrent neural network (RNN) [19], bidirectional recurrent neural network (BiRNN) [20], long short-term memory network (LSTM) [21] and gated recurrent unit (GRU) [22]. Gao et al. [23] systematically evaluated the performance of deep learning features in view-based 3D model retrieval on four popular datasets. Li et al. [24] proposed a CNN model to detect, recognize and abstract the information in the input road scene, then a decision-making system calculates the specific commands to control the vehicles. Nevertheless, CNN extracts features without time correlation. In this way, the features extracted from network for decision-making limit the accuracy rate of driving decisionmaking. To introduce time correlation, a RNN employed for driving decision making. Chen et al. [25] pretended a novel model for self-driving cars which comprises a CNN, a cognitive map and a RNN to learn the driving behavior and decision-making process of a human driver. However, a RNN is unable to resolve long-term dependencies. To address longterm dependencies in decision making, LSTM adapted in autonomous driving. Chen et al. [26] developed a deep CNN-LSTM algorithm for self-driving simulation to control the movement of self-driving vehicles in the driving simulation. However, the memory ability of important features and classification capability of LSTM is poor.
Softmax classifier is used in most deep learning methods. However, the poor classification ability of the softmax classifier leads to poor classification results. Support vector machine (SVM) [27] is introduced because of its stronger classification ability. The classification results of SVM are largely affected by the parameters of kernel function. Metaheuristic optimization algorithm is applied to optimize the parameters of kernel function.
For metaheuristic optimization algorithm, Kennedy and Eberhart [28] proposed Particle Swarm Optimization (PSO), which is inspired by the act of searching for food, each member of the swarm continuously changes its search pattern by learning from its own experience and the experience of other members. Boettcher and Percus [29] proposed extremal optimization (EO), its inspiration is self-organized criticality, a concept introduced to describe emergent complexity in physical systems. Hashim et al. [30] proposed a novel metaheuristic algorithm named Henry gas solubility optimization (HGSO), which mimics the behavior governed by Henry's law to solve challenging optimization problems. Heidari et al. [31] proposed Harris hawks optimization (HHO), its main inspiration is the cooperative behavior and chasing style of Harris' hawks in nature. Li et al. [32] proposed slime mould algorithm (SMA) based on the oscillation mode of slime mould in nature. Faramarzi et al. [33] proposed marine predators algorithm (MPA), which is inspired by the widespread foraging strategy along with optimal encounter rate policy in biological interaction between predator and prey. Saremi et al. [34] proposed grasshopper optimization algorithm (GOA), which is a new nature-inspired algorithm designed by grasshopper swarm characteristics. There are only two parameters of kernel function in SVM that need to be optimized, GOA is chosen because of strong stability and simpler optimization process.
However, exploration and exploitation capabilities of GOA are not coordinated. Luo et al. [35] proposed an improved GOA which combines three strategies including Gaussian mutation, Levy-flight strategy and opposition-based learning. Sulaiman [36] updated the GOA with a better initialization strategy to balance the search capability. Zhao et al. [37] proposed improved GOA with nonlinear comfort zone parameter, the Levy flight mechanism and random jumping strategy. Raeesi et al. [38] improved GOA by adding opposition-based learning and merit function methods to boost its exploration and exploitation abilities.
To sum up, three issues are not resolved in the above methods based on deep learning. First, current algorithms have simple network structure and not consider the information of the surrounding vehicles, which leads to inadequate factors considered in the decision of the ego vehicle and inaccurate decision results. Second, softmax classifier is used in existing algorithms, which leads to their poor classification accuracy. Third, fixed kernel function parameters lead to poor classification ability in SVM classifier after replaced softmax classifier. Moreover, in the process of using GOA to optimize parameters of SVM, exploration and exploitation capabilities of GOA are not coordinated.
To solve above issues, we propose a novel network architecture named improved LSTM network with SVM classifier optimized by GOA (GOA-ImLSTM). Firstly, in order to consider information of surrounding vehicles, a novel network architecture called improved LSTM (Im-LSTM), used to extract vital features for self-driving vehicles, with three parallel LSTM network units and a network unit serial connected according to vehicle location is constructed. The first three parallel LSTM network units extract the key features containing vital status information in the left lane, the right lane, and the front vehicle that facilitate the ego vehicle to understand surrounding environment and make correct driving decision-making. Then the extracted features are concatenated with the state of the ego vehicle, and feed it to the last LSTM network unit for obtaining the driving decision-making. Secondly, to enhance the decision-making classification accuracy for self-driving vehicles, SVM with stronger classification capability than softmax classification is employed. Softmax classifier simply maps input vectors to categories, which results in low classification accuracy under complex samples, while SVM classifier with radial basis function kernel maps the samples to a higher dimensional space, which more appropriate to classification under complex samples. In this way, the increasing accuracy of driving decision-making is ensured. Thirdly, since the performance of SVM is affected by the parameters in the kernel function, GOA is used to optimize the parameters of SVM. However, GOA not take into account that the found food source is locally optimal in the early stage of the GOA and the found food source gradually tended to be globally optimal in the later stage of the GOA. This leads to uncoordinated exploration and exploitation capabilities of GOA. Thus, dynamic weights in position movement formula are defined to balance the effects of the interaction of grasshoppers and tendency towards food source on grasshoppers position during the optimization process. A dynamic weight that is increased with the number of iterations for tendency towards food source is defined, and a dynamic weight that is decreased with the number of iterations for interaction of grasshoppers is defined in position movement formula. In this way, in the early stage of the GOA, the position of grasshoppers was more affected by the interaction of grasshoppers. And in the later stage of the GOA, the position of the grasshopper was more affected by the tendency to move towards the source of food, that is more suitable to pattern of grasshoppers motion. Then GOA is easier to find excellent parameters which increase classification accuracy for SVM.
The rest of our paper is organized as follows. In section II, we introduce the related work of our research. In section III, we propose a network architecture called GOA-ImLSTM. In section IV, the whole procedure of the GOA-ImLSTM is introduced. In section V, we evaluate the performance of proposed algorithm with three improvements on Next Generation SIMulation (NGSIM) [39]. In section VI, we give our conclusion.

II. RELATED WORK
In this section, we introduce a LSTM network, a classifier named SVM and a GOA.

A. LONG SHORT-TERM MEMORY NETWORK
Hochreiter et al. proposed LSTM network in 1997 [21]. LSTM network is a special type of RNN that can learn long-term dependence information. In the structure of LSTM network, LSTM cell replaced traditional RNN hidden cell to construct LSTM network. LSTM network has been widely used in speech recognition, language modeling, sentiment analysis, and text prediction. The structure of LSTM cell is shown in Fig. 1. As shown in Fig. 1, the LSTM cell includes three calculation gates, forget gate, input gate and output gate. Each gate is responsible for different function, the forget gate is responsible for determining to keep the unit status of previous moment to the unit status of current moment, the input gate is responsible for determining to keep the input to the unit status of current time, the output gate is responsible for determining the output of the unit status at the current moment. The computational process in LSTM cell is presented as follows.
where x t is input data of current moment, h t−1 is output data of previous LSTM cell, w f , w i , w c , w o are weights of input data x t and previous output data o t are the output of forget gate, input gate, generated information, and output gate. c t−1 and c t are the unit status of the previous moment and current moment. h t is the output data of LSTM cell. VOLUME 8, 2020 σ and tanh are represented the sigmoid and tanh function as follows. [27] is one of the most robust and accurate methods in all well-known classification algorithms. SVM is a supervised learning method that can be widely used in statistical classification and regression analysis. The basic model of SVM is defined as the linear classifier with the largest interval in the feature space. The purpose of SVM is to find an optimal segmentation surface to maximize the classification interval.
Only considered a two-classification problem, the input point is represented by x i ∈ R d (i = 1, 2, · · ·, N ), which is Ndimensional vector, and the category is represented by y i ∈ {−1, 1} (i = 1, 2, · · ·, N ), which represented two different classes. Assume that two classes are linearly separable. The purpose of classifier is to find a hyperplane in N -dimensional data space, which is defined as follows.
To find the optimal hyperplane that maximizes the distance between the closest training sample and the hyperplane. By simply rescaling the hyperplane parameters, the distance can be calculated as equal to 1/ ω , the problem of maximizes distance transforms the minimization problem of ω . Furthermore, the minimization problem of ω is equivalent to as follows.
   minimize : This classic linearly constrained optimization problem can be transformed into the following optimization problem of dual variables by Lagrangian formulation as follows.
where α i is Lagrange multiplier which is estimated using quadratic programming methods in [40]. Assume the solution to the optimization problem of dual variables has been obtained as follows.
The ω and b of optimal hyperplane are defined as follows.
The optimal hyperplane is redefined as follows.
The classification decision function is defined as follows.
where sign is a symbolic function, which is defined as follows.
proposed GOA in 2017 [34]. GOA is a new nature-inspired algorithm designed by two grasshopper swarm characteristics. First characteristic is slow movement and small steps of the grasshoppers in the larval phase, as well as long-range and abrupt movement in adulthood. Second characteristic is food source seeking of grasshoppers. Two characteristics of grasshoppers performed two tendencies of search process in nature-inspired algorithms, which are exploration and exploitation, as well as target seeking.
To model the swarming behaviors of grasshoppers, the mathematical model is presented as follows.
where X i defines the position of the i-th grasshopper, S i is the social interaction, G i is the gravity force on the i-th grasshopper, and A i shows the wind advection. S i presented the social interaction is defined as follows.
where d ij is the distance between the i-th and the j-th grasshoppers, s is a function to define the strength of social forces, is calculated as follows.
where f indicates the intensity of attraction and l is the attractive length scale, we have chosen l equals 1.5 and f equals 0.5 in [34]. And d ij is a unit vector from the i-th grasshopper to the j-th grasshopper as follows.
G i is calculated as follows.
where g is the gravitational constant and e g shows a unity vector towards the center of the earth. A i is calculated as follows.
where u is a constant drift and e w is a unity vector in the direction of wind. However, this mathematical model cannot be used directly to solve optimization problems, a modified equation is proposed to solve optimization problems in [34]. (24) where ub d is the upper bound in the Dth dimension, lb d is the lower bound in the Dth dimension, T d is the value of the Dth dimension in the target, and c is a decreasing coefficient to shrink the comfort zone, repulsion zone, and attraction zone.
To balance exploration and exploitation in search process, the parameter c is required to be decreased proportionally to the number of iterations. Decreasing coefficient c is defined as follows.
where cmax is the maximum value, cmin is the minimum value, l indicates the current iteration, and L is the maximum number of iterations. We have chosen 1 and 0.00001 for cmax and cmin in [34].

III. GOA-IMLSTM
GOA-ImLSTM is proposed based on LSTM, SVM, and GOA as shown in Fig. 2. Three improvements are presented in this section. In part A, a novel network architecture with three parallel LSTM units and a LSTM unit serial connected according to vehicle location is designed, termed Im-LSTM. In part B, the commonly used softmax is replaced by SVM to accomplish the classification task. In part C, improved GOA is applied to optimize the parameters of the SVM, termed ImGOA-SVM.

A. IM-LSTM
Im-LSTM consists of four LSTM network units, which are Lstm_left, Lstm_right, Lstm_front and Lstm_ego. The input of Im-LSTM contains the status of six vehicles as shown in Fig. 3, including angle steer, speed, acceleration, lateral distance from ego vehicle, longitudinal distance from ego vehicle, vehicle length, vehicle width. Because driving decisions are time-dependent, the status of each vehicle in the past T time steps is selected as input. To consider the information of the surrounding vehicles, network architecture, used to extract important features for self-driving vehicle, with three parallel LSTM units and a LSTM unit serial connected according to vehicle location is designed. As shown in Fig. 2, Im-LSTM first splits the information of the surrounding vehicles by vehicle location and feeds them  weight of front vehicle. And the sum of three weights is 1, the α, β, and γ are set as 0.4, 0.4 and 0.2.

B. SVM CLASSIFIER
Softmax classifier is used in most methods based on LSTM for decision-making. However, the classification ability of softmax classifier is poor, which leads to inaccurate classification results. To solve this issue, SVM classifier is employed for classifying the driving decision-making in GOA-ImLSTM.
The input of SVM classifier is the key feature for driving decisions from Im-LSTM, and the output of SVM classifier is the final decision-making of ego vehicle. SVM classifier in GOA-ImLSTM as shown in Fig. 4. The kernel function of SVM classifier chose radial basis function (RBF) kernel. RBF kernel function is chosen because the fitting ability of RBF kernel function is better than logistic regression classifier by mapping samples to a higher dimensional space. The classification ability of SVM classifier with RBF kernel function depends heavily on the choice of c and γ , the parameters of RBF kernel function. Therefore, the main purpose of the next part is to optimize c and γ of RBF kernel function.

C. IMGOA-SVM
To achieve excellent classification results, the parameters c and γ of RBF kernel function obtained through optimization algorithm. The parameter c is the penalty coefficient, which is the tolerance of the error. The larger c indicates a lower tolerance for errors, which easily leads to overfitting. The smaller c is more easily to cause underfitting. The parameters γ implicitly determine the distribution of the data after it is mapped to the new feature space. Larger γ results in fewer support vectors, and smaller γ results in more support vectors. The number of support vectors affects the speed of the training and prediction process in classification. Consequently, the values of parameters c and γ are very important to the accuracy of classification results.
In order to promote the classification ability of SVM, the GOA method is introduced to optimize the parameters c and γ of SVM, termed GOA-SVM. Optimizing parameters of SVM by using GOA includes three aspects. First, each grasshopper means a set of solutions for c and γ . Thus, the movement of grasshopper represents the changes in parameters c and γ , and the searching of the best grasshopper is to find the best c and γ . The performance of each grasshopper is evaluated by the fitness function. The parameters c and γ are updated based on the fitness values.
Furthermore, the movement equation of grasshoppers in grasshopper optimization algorithm is defined in (24). The first term of movement equation considers the position of other grasshoppers and implements the interaction of grasshoppers in nature. The second term of movement equation simulates their tendency to move towards the source of food. Nevertheless, in the early stage of the grasshopper optimization algorithm, the found food source was locally optimal, thus, the position of grasshoppers was more affected by the interaction of grasshoppers. In the later stage of the grasshopper optimization algorithm, the found food source gradually tended to be globally optimal, so the position of the grasshopper was more affected by the tendency to move towards the source of food. To balance interaction of grasshoppers and tendency towards food source, dynamic weights in position movement formula are defined.
At the beginning of the optimization process, the interaction of the grasshoppers has a great influence on the position of the grasshoppers. As the number of iterations increases, the positions of the grasshoppers gradually move to the source of food, which is less affected by the interaction of the grasshoppers. Thus, a gradually decreasing dynamic weight is defined for the interaction of the grasshoppers. And at the beginning of the optimization process, the tendency to move towards the source of food has little effect on the grasshoppers position, since the current source of food is locally optimal. As the number of iterations increases, the influence of the tendency to move towards the source of food becomes greater, since the current source of food gradually tends to global optimization. Thus, a gradually increasing dynamic weight is defined for the tendency to move towards the source of food.
A modified equation is proposed as follows. (26) where w 1 is the weight of interaction of grasshoppers, w 2 is the weight of tendency towards food source. The linearly decreasing and increasing weights leads unbalanced exploration and exploitation capabilities of GOA, to match these two capabilities and enhance the search ability of the algorithm, the exponential and logarithmic functions with good performance are introduced to nonlinear adjust parameters. w 1 and w 2 are defined as follows.
155434 VOLUME 8, 2020 where l indicates the current iteration, and L is the maximum number of iterations. The graph of weights w 1 and w 2 is presented in Fig. 5 and Fig. 6.
As shown in Fig. 5, w 1 decreases as the number of iterations increases, when the number of iterations is extremely small, w 1 is close to 1, at this time the position of grasshoppers was more affected by the interaction of grasshoppers.
In contrast, as shown in Fig. 6, w 2 increases as the number of iterations increases, when the number of iterations is extremely high, w 2 is close to 1, at this time the position of grasshoppers was more affected by the tendency towards the source of food.
Improved GOA with new position formula is employed to optimize parameters of SVM, termed ImGOA-SVM.

IV. THE OVERALL PROCESS OF GOA-IMLSTM
The flowchart of GOA-ImLSTM is illustrated in Fig. 7. First, Im-LSTM trains the input samples and obtain the key feature for driving decision-making. Second, the positions of N grasshoppers, which presented parameters of SVM, are randomly initialized at the beginning of ImGOA-SVM. Then, the fitness values of grasshoppers are calculated by obtaining the accuracy of SVM classification results with RBF kernel of parameters c and γ . Next, update the position of all grasshoppers in the population according to the defined movement Algorithm 1 GOA-ImLSTM Input: S: status of surrounding vehicles and ego vehicle Output: D: decision-making of ego vehicle 1 split S to S l , S r , S f , S e by vehicle location 2 obtain pivotal features of left lane F l by feeding S l to Lstm_left 3 obtain pivotal features of right lane F r by feeding S r to Lstm_right 4 obtain pivotal features of front lane F f by feeding S f to Lstm_front 5 F ← weight F l , F r , F f separately and add 6 F ← F concatenated with S e 7 obtain pivotal features F e by feeding F to Lstm_ego 8 iter ← 1 9 while iter < Max_iterations do 10 Initialize the positions of the grasshoppers 11 Calculate the fitness of grasshoppers by SVM 12 Update the food source 13 iter ← iter + 1 14 Update the positions of grasshoppers by (26) 15 D ← classification by SVM with parameters by optimal position equation in (26). Furthermore, if the maximum iteration of GOA is reached, the optimization process is finished. As a result, we can obtain the optimal parameters c and γ . The searching range for parameters c and γ is [0.01, 10] and [0.0001, 5] respectively.

V. EXPERIMENT AND ANALYSIS
In order to evaluate the performance of the GOA-ImLSTM, we have applied the algorithm on NGSIM. The data processing platform is windows10 on laptop PC 2.3GHz processor and 4GB memory.

A. DATASET AND PROCESSING
Researchers at the NGSIM project collected detailed vehicle trajectory data using data from US 101 Southbound Road from Los Angeles, California [39]. Vehicle trajectory data provides the precise position of each vehicle in the study area every tenth of a second to obtain detailed lane positions and positions relative to other vehicles. We select representative data for training and testing from dataset.
First, find the turning angle of each vehicle at the current moment by formula as follows.  Second, the dataset is grouped by time to complete the status information of each vehicle, including the leftfront vehicle, the left-behind vehicle, the right-front vehicle,

B. PARAMETERS SETTING
The parameters setting in GOA-ImLSTM is introduced in table. 1. All the codes are implemented in Tensorflow 1.12.0 and Keras 1.3.0., and the network was trained by the Adam optimizer.

C. EXPERIMENT RESULTS AND ANALYSIS
In this section, we evaluate the performance of the proposed network architecture on NGSIM and compare the results with other methods. The network has 48,000 training samples and 3,600 testing samples, and each decision has the same number of samples.
As shown in Fig. 10, the accuracy of driving decisionmaking increases with the number of iterations. The maximum number of iterations of GOA-ImLSTM is 200. When the number of iterations is between 1 and 50, the decision accuracy increases with the number of iterations. If speed is required but accuracy is not required, the number of iterations is selected to be 20. If speed and accuracy are both required, the number of iterations is selected to be 50. When the number of iterations is between 50 and 100, the rate of decision accuracy increases decreases. If speed is required but accuracy is more required, the number of iterations is selected to be 100. When the number of iterations is between 100 and 200, the training time is very long but the increase in accuracy slows down significantly, so we not consider the case of 200 iterations. We compared the driving decision-making task against six methods: RNN, LSTM, GRU, Im-LSTM, Im-LSTM with SVM and Im-LSTM with GOA-SVM, which network architectures as shown in table. 2.  Table. 3, 4, and 5 shows the quantitative results of GOA-ImLSTM on the NGSIM dataset. The criterion is the accuracy rate. The results show that three innovations gradually  improve the accuracy of decision results on the NGSIM dataset.
As shown in table. 3, 4, and 5, the performances of RNN, GRU, and LSTM are compared in first, second and third lines. LSTM has the highest overall accuracy rate in the case of 20 iterations, 50 iterations and 100 iterations. That proves LSTM network structure is promising in driving decision-making. The advantage of LSTM compared with RNN is time dependence, and compared with GRU, LSTM has more accurate decision-making in this case of large data volume because of its more parameters. And as the number of iterations increases, the accuracy rate of decision-making increases.
The comparison of LSTM and Im-LSTM is shown in third and fourth lines in table. 3, 4, and 5. For overall accuracy rate, Im-LSTM is 30.3% better than LSTM in the case of 20 iterations, 23.3% better in the case of 50 iterations and 23.1% better in the case of 100 iterations. That indicates the novel network architecture of Im-LSTM indeed promotes the accuracy rate of self-driving decision-making because Im-LSTM extracts more accurately characteristics of the situation at that time to make more accurate decisions.
Fourth and fifth lines in table. 3, 4, and 5, the performances of Im-LSTM and Im-LSTM with SVM are compared. In the case of 20 iterations, Im-LSTM has higher accuracy rate in turn right lane decision-making, but Im-LSTM with SVM has higher lane-keeping decision-making, turn left lane decisionmaking and overall accuracy rate. Although the accuracy rate of Im-LSTM and Im-LSTM with SVM network in some specific decision-makings are equal, Im-LSTM with SVM  has higher overall accuracy rate in the cases of 50 and 100 iterations. According to this comparison, the SVM classifier indeed increases decision precision because its classification ability is stronger than softmax classifier.
The comparison of Im-LSTM with SVM and Im-LSTM with GOA-SVM is shown in fifth and sixth lines in table. 3, 4, and 5. Although these two methods have equal accuracy rate in lane-keeping in the case of 20 iterations, Im-LSTM with GOA-SVM has higher accuracy rate in turn left lane, turn right lane decision-making and overall. For overall accuracy rate, Im-LSTM with GOA-SVM is 0.8% better than Im-LSTM with SVM in the case of 20 iterations, 2.1% better in the case of 50 iterations and 2.1% better in the case of 100 iterations. The comparison shows that GOA finds more suitable parameters for the classification of driving decisionmaking to improve the classification ability of SVM, so Im-LSTM with GOA-SVM obtained a higher accuracy rate.
Last two lines in table. 3, 4, and 5, the performances of Im-LSTM with GOA-SVM and GOA-ImLSTM are compared. The overall accuracy rate of GOA-ImLSTM has 2.6% improvement over LSTM with GOA-SVM in the case of 20 iterations, 2.1% improvement in the case of 50 iterations and 2.3% improvement in the case of 100 iterations. That demonstrates the improved GOA gets more excellent parameters of SVM due to the balance of exploration and exploitation ability, thereby further improving the performance of SVM classifier. Table. 6 shows the comparison of GOA-ImLSTM and the previous state-of-the-art methods, including lane change intent prediction method [41], game-theoretical decisionmaking model [42], bidirectional recurrent neural network [43], bayesian network [44], J48-Decision tree [44], free lane-changing model [44], and a novel approach using a combination of spectral graph analysis and deep learning [45]. It shows that the overall accuracy boost to 93.0%, which has 1.6% improvement over lane change intent prediction method, 7.0% improvement over game-theoretical decision-making model, 12.4% improvement over bidirectional recurrent neural network, 21.4% improvement over bayesian network, 17.6% improvement over J48-Decision tree, 6.1% improvement over free lane-changing model, and 1.8% improvement over a novel approach using a combination of spectral graph analysis and deep learning. The results clearly show that GOA-ImLSTM achieves the best performance in terms of overall accuracy rate.

D. DISCUSSION
According to above observations, it can be concluded that the performance of GOA-ImLSTM is better than other methods in autonomous driving decision-making. The reason is that each innovation contributes to the improvement of decisionmaking accuracy. The runtime of GOA-ImLSTM is 520µs in driving decision-making.
Although GOA-ImLSTM has obtained good experimental results, it still has certain limitations. First, the computational complexity of the algorithm is relatively high, which is O(n 2 ). Second, the final decision classification result still largely depends on the parameters of the SVM classifier. Despite we use an improved GOA to optimize the parameters in SVM, since the computational complexity cannot be too high and the number of optimization iterations is limited, there is no guarantee that the parameters after a fixed number of optimizations will be the best results.

VI. CONCLUSION
Because the information of the surrounding vehicles is not considered and poor classification capabilities of LSTM, the classification accuracy rate of decision-making by using LSTM is unsatisfactory. To solve this issue, we presented an efficient network architecture for driving decision-making which named GOA-ImLSTM. First, we designed a novel network architecture with three parallel LSTM units and a LSTM unit serial connected according to vehicle location to consider the information of the surrounding vehicles for making decisions. Second, we introduced SVM classifier instead of universally used softmax to accomplish the classification task because of its stronger classification capability. Third, we employed GOA to optimize the parameters in SVM. Furthermore, we defined dynamic weights in position movement formula. In this way, the optimization effect was further enhanced.
The results verify that the GOA-ImLSTM achieves better performance on the NGSIM dataset compared with other methods. In the future, we will apply the proposed network architecture to other tasks and continue to improve.