Person-vehicle-road cooperative unmanned driving credibility acceleration evaluation method

Abstract

The invention provides a man-vehicle-road cooperation unmanned driving credibility acceleration evaluation method, which comprises the following steps: generating a test scene case based on man-vehicle-road cooperation unmanned driving data and an unmanned driving typical operation scene, and building an unmanned driving-oriented black box system test platform; then, single test scene evaluation is realized, multiple classic unmanned driving black box algorithms are trained, then, system acceleration test evaluation is performed, a sample set is iteratively updated by using subset simulation in combination with an adaptive Markov chain Monte Carlo method, and acceleration test evaluation of an unmanned driving failure scene is realized; and finally, carrying out acceleration effect evaluation and optimization from the angles of algorithm effectiveness and algorithm efficiency, and outputting an acceleration evaluation result oriented to human-vehicle-road cooperation unmanned driving credibility evaluation. The method is suitable for the unmanned driving algorithm with the black box characteristic, the evaluation efficiency is improved by 20-200 times, and the method has the advantages of being easy to expand, capable of being generalized, suitable for various typical unmanned driving evaluation tasks and the like.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving testing and evaluation, and in particular to a method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving. Background Art

[0002] With the development of driverless technology, highly automated vehicles are applied in actual traffic systems to provide more convenient and safe travel options. In recent years, black box algorithms based on deep learning and reinforcement learning have made significant progress in the application of driverless driving. Although end-to-end black box algorithms have stronger autonomous learning capabilities and high-dimensional data processing capabilities compared to traditional algorithms based on predefined rules, poor scalability and efficiency, they sacrifice their verifiability and may cause unpredictable failures in long-tail scenarios.

[0003] Due to the huge harm caused by the effectiveness of the algorithm, it is crucial to quantify and evaluate the safety performance of the black box algorithm for autonomous driving in a human-vehicle-road collaborative environment. The performance of autonomous vehicles in boundary scenarios is an important manifestation of their credibility. However, the current evaluation methods for autonomous driving black box algorithms are still inefficient, becoming a key obstacle to the development and deployment of autonomous driving systems. In high-dimensional space, the essence of verifying the safety performance of an algorithm is a rare event estimation problem. According to a 2021 report from California, the failure rate of the most advanced autonomous vehicles has reached about 2.10 per mile. -5 , algorithm failure becomes a rare event with low probability. The key to solving the problem of autonomous driving credibility evaluation lies in solving the "curse of dimensionality" caused by time and space complexity and the "sparseness problem" caused by the low proportion of useful information. According to the latest statistical data, autonomous vehicles need to be tested for at least 11 billion miles to clearly prove the safety of autonomous driving. Therefore, the traditional mileage-based real vehicle testing method has a long cycle, low efficiency and high cost. This inspires us to find more critical test scenarios in the simulation environment to improve test efficiency.

[0004] From a mathematical point of view, the solutions to the evaluation problem can be roughly divided into three categories: verification, falsification, and probability estimation; verification is characterized by its rigor, which proves theoretically that the system does not have errors. However, at present, when most unmanned driving manufacturers use end-to-end black box algorithms, this method is not applicable; the falsification method aims to find all failure scenarios from the parameter space, however, this method can provide less information about the test system and is only used to confirm whether the system meets the requirements. When the system is running in an unstructured environment, it is not enough to simply confirm whether there is a failure. Research points out that in this case, it is more important to understand the overall frequency of failures and the relative likelihood of different failure modes, that is, the probability estimation method, which is usually more portable and scalable.

[0005] In summary, autonomous driving credibility evaluation is a prerequisite for the commercial application of autonomous driving. However, autonomous driving testing faces challenges such as the curse of dimensionality, the curse of sparsity, and black box characteristics. How to quickly test and evaluate autonomous driving black box algorithms in complex traffic environments with human-vehicle-road collaboration is a difficult problem that needs to be overcome in the development and commercial application of autonomous driving. Summary of the invention

[0006] In view of the defects of the prior art in practical applications, the present invention provides a method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving.

[0007] The process of the present invention is as follows Figure 1 As shown, it includes the construction of typical test scenarios for unmanned driving, evaluation of single test scenarios, evaluation of system acceleration tests, and evaluation and tuning of acceleration effects. The specific technical solutions are as follows:

[0008] (1) Construction of typical test scenarios for autonomous driving

[0009] Based on real data sets (real vehicle operation data sets), the present invention builds typical application scenarios of unmanned driving, such as following a vehicle, changing lanes, mixed traffic flow, etc.; and constructs a test platform for an unmanned driving black box system, including parsing scene composition, mining scene parameters, extracting vehicle-specific behaviors, and fitting basic parameter distributions to generate test scenario use cases.

[0010] 1) Black box system test platform for unmanned driving

[0011] a. Analyze scene composition

[0012] Based on the selected typical application scenarios of unmanned driving, the scenario elements are analyzed, and a seven-layer scenario element model is considered, namely 1-road structure; 2-road facilities; 3-temporary changes in roads and facilities; 4-traffic participants; 5-climate environment; 6-communication status; 7-self-vehicle status and behavior. The scenario composition is analyzed from functional scenarios to abstract scenarios.

[0013] b. Mining scene parameters

[0014] Based on the abstract scene, we further explore the scene parameters, construct the logical scene, replace the scene elements with parameters that vary within a certain range, describe certain influencing factors in the scene through parameter range and distribution, and realize parametric expression. This provides a basis for using fixed parameters to accurately describe a specific scene or event sequence in the subsequent steps.

[0015] c. Extract vehicle-specific behaviors

[0016] Specifically, data preprocessing is performed based on the actual vehicle operation data set to extract the vehicle's driving behavior in specific application scenarios as the data basis.

[0017] 2) Construction of parameter distribution of typical operation scenarios

[0018] a. Fitting parameters basic distribution

[0019] Based on the extracted vehicle driving data and scene parameter division, combined with distribution fitting or alternative model knowledge, the basic distribution of parameters is fitted to obtain the test parameter distribution function and probability density function for typical test scenarios. Common distribution forms include multi-dimensional Gaussian distribution, Poisson distribution, gamma distribution, Weibull distribution, etc.

[0020] b. Generate test scenario use cases

[0021] Based on the obtained basic distribution and its probability density function, specific test scenarios can be generated through distribution sampling technology as input variables for subsequent tests.

[0022] (2) Single test scenario evaluation

[0023] The present invention takes into account the widespread application of end-to-end black box algorithms in the commercialization of unmanned driving, uses the industry's advanced deep reinforcement learning algorithm to replace the unmanned driving black box algorithm as the evaluation object, and designs a reward function in combination with the operation requirements of the unmanned driving system in a specific application scenario, so that the algorithm converges and can better complete the unmanned driving operation task; further, in combination with the unmanned driving credibility evaluation requirements, the unmanned driving black box algorithm failure evaluation function and failure judgment threshold are designed. Finally, a single test scenario evaluation is achieved: the unmanned driving black box algorithm is connected to the test platform, its simulation data over a period of time is output, its failure evaluation function value is calculated, and it is compared with the failure judgment threshold, so that the evaluation result of the unmanned driving in this specific scenario is divided into "failure" or "safety".

[0024] 1) Autonomous driving black box algorithm training

[0025] Deep reinforcement learning is a popular algorithm for commercial applications of unmanned driving. Compared with traditional algorithms, it has stronger high-dimensional data processing capabilities and autonomous learning capabilities. The core purpose of reinforcement learning is to learn the optimal strategy by trying different behaviors so that the agent can obtain the maximum cumulative reward. In reinforcement learning, at time t, the agent will observe the environment state s t , and according to its existing strategy Π(a t |s t ) Select action a from state space A t At the same time, the model system will give the agent a reward r t , and then advance the state to s t+1This process is repeated in each training until the end of the training. In this process, the main goal of the agent is to obtain the maximum cumulative reward, which is formulated as follows, where the discount factor α∈(0,1].

[0026]

[0027] Designing an appropriate reward function based on the needs of unmanned driving operations can improve algorithm performance and thus more accurately simulate actual commercial algorithms. For example, in the decision-making algorithm of unmanned driving, the three dimensions of safety, efficiency, and comfort can be integrated; the design can be combined with indicators such as TTC (Time to Collision), headway time, and acceleration change rate.

[0028] The data extracted in step 1 is used to train the driverless black box algorithm, and the training goal is to achieve high scores and convergence for each model.

[0029] 2) Failure assessment function design

[0030] In order to reliably evaluate the performance of the unmanned driving black box algorithm in the human-vehicle-road collaborative environment, a credibility evaluation index system is introduced. Specifically, the unmanned driving credibility principles include safety, efficiency, comfort, compliance, etc., which can specifically take into account TTC, collision frequency, operation delays, acceleration change rate, etc.

[0031] Based on the above trustworthy evaluation indicators, a function is further designed to measure the performance of the algorithm in a specific parameter scenario. After the simulation is completed, the function value is calculated based on the simulation output data and compared with the preset failure judgment threshold to determine the performance score of the algorithm in this scenario and whether it fails.

[0032] (3) System Acceleration Test Evaluation

[0033] The present invention adopts a method based on probability estimation to accelerate the evaluation of the failure probability of the unmanned driving black box algorithm in a human-vehicle-road collaborative environment in a simulation environment. Specifically, a SubSet Simulation method (hereinafter referred to as SuS) is proposed in combination with an Adaptive Markov chain Monte Carlo method (hereinafter referred to as ApMCMC), and the latter is used for intermediate process sampling. By iteratively updating the sampled samples, the initial high-scoring test samples can be quickly converted to low-scoring test samples. The high-dimensional problem is converted into multiple low-dimensional sub-problems through an iterative method to reduce the computational complexity, and the failure cause analysis is further realized based on the iterative process.

[0034] 1) Subset simulation combined with adaptive Markov chain Monte Carlo method

[0035] SuS is an adaptive Monte Carlo method that can be used to estimate the probability of rare failures in high-dimensional problems. It reduces the computational complexity by iteratively converting high-dimensional problems into multiple low-dimensional sub-problems. In the implementation of the SuS algorithm, each sub-problem needs to generate conditional samples under given conditions, which can be achieved using the MCMC algorithm. The specific algorithm principle is summarized as follows:

[0036] Performing SuS requires converting the base distribution to a multidimensional standard normal distribution via a one-to-one mapping U = T(X). If the joint probability density function of X is known, the mapping should be defined using the Rosenblatt transformation, and if only the marginal distributions and correlations are known, the joint PDF should be modeled using the Nataf distribution.

[0037] In SuS, a rare failure event that satisfies the condition P0(f(X)<γ) is defined as event F. The basic concept of SuS is to represent event F as M intermediate events F j The intersection of:

[0038]

[0039] Intermediate events are nested within each other, which means:

[0040]

[0041] The final failure probability is estimated as the product of the conditional probabilities:

[0042]

[0043] By selecting intermediate events so that the conditional probability P r (F j |F j-1 ) is as large as possible, avoiding the need to deal directly with rare event estimates. The original problem of estimating the probability of rare failures is reduced to a series of intermediate problems that correspond to the estimation of larger conditional probabilities, starting from the higher probability level where the sub-events are located and gradually approaching the lower probability level where the target event is located.

[0044] Specifically, the threshold of each level is adjusted by scaling to increase the probability of intermediate events. Let the threshold of the j-th intermediate event be γ j , F j It can be expressed as:

[0045] F j ={X∈P0:P0(f*(X)<γ j )} (5)

[0046] γ j Satisfies the following relationship

[0047] γ0>γ1...>γ M =γ (6)

[0048] For each intermediate event, γ j The value of needs to be adaptively selected so that the conditional probability P r (F j |F j-1 ) corresponds to the selected value p0. This step is performed by continuously sampling each conditional probability density function To achieve: that is, generate n samples {u (k) j-1 : k = 1, ..., N} And γ j Set to sample u j-1 The p0 percentile of the f(x) values.

[0049] For the sub-samples in the iterative process, the MCMC method is used to analyze the samples. Sampling is performed by using the subsets that fall into F at level j-1 j-1 The sample {u (k) j-1 : k = 1, ..., N s} as the starting seed to simulate the state of the Markov chain. Since all seeds are based on distributed, so the chains have already reached their steady state at the beginning and no aging period is needed (a transition period is required before a Markov chain approximately reaches its steady state).

[0050] Based on the MCMC algorithm, the ApMCMC algorithm is proposed. The ApMCMC algorithm mainly introduces a scaling parameter β, which is updated according to the difference between the average acceptance rate and the target acceptance rate. The idea of ​​the adaptive process is to perform the simulation in steps. In each step, the conditional sampling algorithm with the same parameter β is applied to simulate N s The fraction of chains (parameter N a is defined as the update frequency of the scaling parameter). For the next N a The simulation of a chain adjusts the parameter β based on the estimated acceptance probability of the previous chain. For each subsequent level, the initial value of the scaling parameter β should be chosen as the final value β obtained in the previous level to ensure that the acceptance rate remains close to the optimal value.

[0051] Repeat the sampling and iteration process until the percentile of p0 is less than or equal to γ, which means that F M and F are already in the same failure level. In this level, the failure probability The formula may be updated as follows, where I represents an indicator function used to determine whether a sample function value is less than a threshold.

[0052]

[0053] u (k) M-1 is from Finally, we can obtain an estimate of the failure probability as follows:

[0054]

[0055] 2) Failure cause analysis

[0056] By observing the change size and direction of each dimension of the sample parameters during the sample iteration process, we can explore the key factors that cause the failure of the unmanned driving black box algorithm, and further analyze the failure causes of the unmanned driving black box algorithm.

[0057] (4) Acceleration effect evaluation and optimization

[0058] The present invention evaluates and optimizes the acceleration effect of the credibility acceleration evaluation algorithm; and comprehensively considers the effectiveness and efficiency of the algorithm to evaluate the performance of the acceleration evaluation algorithm used in (3).

[0059] 1) Evaluation of algorithm effectiveness

[0060] The effectiveness of the algorithm is evaluated by calculating the relative error between the evaluation result of the selected method and the true probability value. First, the failure judgment threshold is selected, and the true value is approximated through a large number of random Monte Carlo simulations. The SuS combined with the ApMCMC algorithm is used to repeatedly estimate the failure probability of the algorithm multiple times (such as 30 times), and the average value is calculated as the final evaluation probability of the algorithm; the relative error between the evaluation probability and the true probability is calculated to evaluate the effectiveness level of the algorithm. The calculation formula is as follows:

[0061]

[0062] 2) Evaluation of algorithm efficiency

[0063] The efficiency of the algorithm can be divided into two main aspects:

[0064] a. Compare the selected method with the traditional Monte Carlo method in terms of estimation accuracy, estimation variance, confidence interval, and computational time.

[0065] For estimation accuracy, the relative errors of the two methods under the same test samples and different failure judgment thresholds are calculated and compared;

[0066] For the estimation variance, the dispersion of data and the stability of the estimated value can be measured. The estimation variance of the two methods under the same sample and different failure judgment thresholds can be calculated and compared.

[0067] The confidence interval can reflect the uncertainty of probability estimation. The confidence interval sizes of the two methods under the same sample and different failure judgment thresholds are calculated and compared.

[0068] The calculation time can reflect the efficiency of the algorithm by calculating the time ratio required for the two methods to achieve the same level of sampling accuracy under different failure judgment thresholds.

[0069] b. Compare the selected method with other rare event simulation algorithms

[0070] In addition to the traditional Monte Carlo method, this method is compared with the SuS algorithm and other rare event simulation algorithms such as importance sampling and sequence importance sampling methods in various dimensions, and the efficiency of this method relative to other methods can be analyzed.

[0071] Compared with the prior art, the trust perception method and system provided by the present invention has the following technical key points and advantages:

[0072] (1) The present invention takes into account the black-box characteristics of the algorithms used in actual commercial unmanned vehicles and provides an end-to-end evaluation method.

[0073] (2) The present invention addresses the curse of dimensionality and the curse of sparsity and achieves accelerated evaluation of unmanned driving black box algorithms.

[0074] (3) The method and system for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving proposed in the present invention are scalable and can be easily transplanted into various typical unmanned driving scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0076] Figure 1 A flow chart of the method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving provided by the present invention;

[0077] Figure 2 A schematic diagram of the vehicle following test scenario elements and their corresponding parameters provided in an embodiment of the present invention;

[0078] Figure 3 An integrated theoretical framework for unmanned driving system evaluation provided for the examples of the present invention;

[0079] Figure 4 It is a schematic diagram of the front vehicle speed dimension in a basic distribution P0 corresponding to the test sample X;

[0080] Figure 5(a) to Figure 5(b)A schematic diagram of experimental verification of the effectiveness of the acceleration evaluation algorithm provided by the example of the present invention;

[0081] Figure 6(a) to Figure 6(b) The first schematic diagram of the experimental verification of the efficiency of the acceleration evaluation algorithm provided by the example of the present invention;

[0082] Figure 7(a) to Figure 7(b) The second schematic diagram of the experimental verification of the efficiency of the acceleration evaluation algorithm provided by the example of the present invention;

[0083] Figure 8 The accelerated evaluation results of the credibility evaluation of the human-vehicle-road collaborative unmanned driving decision-making algorithm provided by the example of the present invention;

[0084] Fig. 9 for Figure 4 Schematic diagram of the initial sample u0 corresponding to the basic distribution P0 for the rear vehicle velocity and the front vehicle acceleration slice. DETAILED DESCRIPTION

[0085] The present invention relates to a method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving. The following figure further illustrates the specific implementation of the present invention in combination with the accompanying drawings and specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be pointed out that for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0086] This embodiment is based on the real data set NGSim-I80. Taking the decision-making part of the unmanned driving system as the "brain" among the three major components of perception, decision-making, and control as an example, this embodiment realizes the accelerated evaluation of the credibility evaluation of the intelligent decision-making algorithm for human-vehicle-road collaborative unmanned driving, carries out the construction of vehicle-following test scenarios, the training of black-box decision algorithms based on deep reinforcement learning, the accelerated test evaluation of decision algorithms, and the algorithm performance tuning, and builds a theoretical system framework for unmanned driving evaluation. Figure 3 As shown, the specific implementation method includes the following steps:

[0087] S1: Construction of typical test scenarios for autonomous driving

[0088] like Figure 2 As shown in the figure, a classic following-car scenario is selected as the test scenario, in which the front vehicle simulates other vehicles in the actual road environment, and the rear vehicle simulates an autonomous driving vehicle, which is controlled by a black box decision algorithm based on deep reinforcement learning.

[0089] The key test parameters of the scene are four-dimensional, including the initial speed LvSpeed0 of the front car, the initial speed FvSpeed0 of the rear car, the initial distance Distance0 between the two cars, and the acceleration a of the front car. Lv.

[0090] During the simulation, the black box decision algorithm of the following vehicle takes the current four-dimensional parameters as input and outputs the acceleration a of the following vehicle at the next time step. RL , and the remaining parameters are changed according to physical formulas.

[0091] Assume that the scenario is updated at intervals of Δt, and a total of i time intervals have passed. The total length of the simulation is ΔT, and the speed difference between the two vehicles in the i-th time interval is relSpd i , so the update formula of the scene parameters is as follows:

[0092]

[0093] When sampling the test scene, the test sample X It can be expressed as:

[0094] X=(Distance0, LvSpeed0, FvSpeed0, a Lv ) (11)

[0095] This step is based on the real data set NGSim-I80. A total of 1341 car-following data and 323739 real-time data on the speed of the front and rear vehicles and the following distance are extracted. The distributed environment of human-vehicle-road collaboration is constructed by the distribution fitting method, and the basic distribution P0 is obtained. The distribution fitting diagram of the front vehicle speed is shown in the figure. Figure 4 As shown in Figure 2, the specific test scenario is obtained by sampling parameters from the distribution.

[0096] S2: Single test scenario evaluation

[0097] In this step, four classic deep reinforcement learning models are first trained to simulate the black box decision algorithm of driverless vehicles as the object for subsequent credibility acceleration test evaluation. The reinforcement learning reward function is designed by comprehensively considering the three major indicators of safety, efficiency, and comfort, as follows:

[0098] 1) Security

[0099] Safety is one of the main training objectives of reinforcement learning, as the most critical factor considered by autonomous vehicles in the subsequent decision-making process. The time to collision (TTC) metric is used to describe safety, and the following decision-making behaviors with TTC values ​​less than 4 seconds should be penalized. TTC is a type of safety surrogate measure (SSM) and is the most common indicator for quantifying the risk of traffic conflicts. It represents the remaining time before two vehicles collide. TTC and safety indicator F TTC The formula is as follows:

[0100]

[0101] In addition, considering the serious negative impact of vehicle collision accidents on safety, a collision penalty term is introduced. The penalty coefficient is represented by λ, which is 100. Whether a collision accident occurs in a single simulation process is represented by the Boolean value judgecollision. If a collision occurs, the value is 1, otherwise it is 0.

[0102] 2) Efficiency

[0103] The headway is an important indicator of road capacity. The efficiency of following a vehicle is reflected in the shorter following distance or headway time under the premise of ensuring safety. The logarithmic distribution of the headway interval of the NGSIM-I80 dataset is constructed, and the probability density function is used as the efficiency indicator. Considering the priority of safety, the efficiency reward is only designed when the safety index is zero. Headway time h t and efficiency index F headway The calculation is as follows:

[0104]

[0105] Among them, formula (14) is a theoretical formula, t represents the current time, that is, the headway at the current time is equal to the distance between the two vehicles at the current time divided by the speed of the rear vehicle at the current time, and the speed of the rear vehicle at the current time FvSpeed t The value is the speed of the following vehicle corresponding to Train in formula (10), that is, the speed of the following vehicle in the real dataset NGSim-I80.

[0106] About formula (15)f lognorm The explanation is as follows: Based on the headway data of 1341 car-following data in the real data set, the lognormal distribution is fitted. The lognormal distribution is a probability distribution whose logarithm is normally distributed. headway According to the current headway h t And the lognormal distribution f lognorm μ, σ are the parameters of exponential normal distribution, and their values ​​are μ=0.42, σ=0.44.

[0107] 3) Comfort

[0108] The acceleration rate is a key indicator that significantly affects passenger comfort. When subjected to a considerable acceleration rate, passengers will feel noticeable discomfort. Using it as an indicator to quantify comfort, the comfort index can be expressed as follows:

[0109]

[0110] The final expression of the reward function is as follows:

[0111] reward=10·FTTC +F headway +F jerk -λ·judgecollision (17)

[0112] During training, the vehicle following data extracted from NGSim-I80 is used as the training data set, so that the RollingScore of the model to be evaluated converges and shows satisfactory performance in the given test scenario.

[0113] Finally, in order to evaluate the safety performance of the intelligent decision-making algorithm in a specific test scenario, a failure evaluation function is designed. In a single test scenario, a safety evaluation is performed on the unmanned driving black box decision algorithm connected to the test scenario to obtain the value of the safety evaluation function. This step can extend the safety evaluation function to other evaluation indicators such as the interaction complexity function. The safety evaluation function is designed as follows:

[0114] Assume that the simulation interval is 0.1s, i.e. 10HZ, and the simulation duration is 10s. The simulation will terminate when the total duration is reached or the two vehicles collide. For the scene samples sampled from the basic distribution, they are input into the simulation environment. After the simulation test starts, the simulated unmanned vehicles interact through the intelligent decision-making algorithm and output the time series of the following scene parameters during the simulation process. The TTC value sequence of the whole process is calculated and its minimum value is set as the safety evaluation function value during the simulation. The safety evaluation function formula is as follows:

[0115]

[0116] At this point, a single test scenario evaluation is achieved, and the black-box decision algorithm is connected to the vehicle-following simulation environment. The specific test scenario is obtained by sampling the basic distribution of parameters fitted in the vehicle-following simulation environment, and input into the black-box decision algorithm and the simulation environment. The black-box decision algorithm generates the acceleration of the following vehicle according to the scenario, and runs the simulation process until it stops according to the pre-defined scenario parameter update formula to obtain the simulation output result.

[0117] The safety assessment function is calculated and compared with the failure judgment threshold. If it is less than, the black box decision algorithm fails in the sampling test scenario, and the scenario is the "failure scenario" of the algorithm; if it is greater than or equal to, the black box decision algorithm makes a safe decision in the sampling test scenario.

[0118] S3: System Acceleration Test Evaluation

[0119] This step uses subset simulation combined with the adaptive Markov chain Monte Carlo method to effectively estimate the failure rate of the driverless black box decision algorithm in the simulation scenario. It can be mainly divided into the following three steps:

[0120] S31: Initialization settings

[0121] 1) Initialization parameters, where the number of samples in each intermediate step is set to N = 1000, the subset threshold probability is set to p0 = 0.1, the number of simulation rounds is set to NSIM = 30, and the MCMC method scaling factor is β.

[0122] 2) Design a distribution transformation function to convert samples from any distribution into samples from a standard normal distribution, and vice versa. Specifically, the basic distribution P0 is first converted into a standard distribution for updating samples. When calculating the value of the security assessment function or generating the final output sample, the converted sample is reconverted into a sample from the original distribution.

[0123] S32: Monte Carlo method to generate initial samples

[0124] Generate an initial sample u0 from the base distribution P0. Use the security evaluation function f(x) to calculate the corresponding function value, denoted as geval0. In addition, calculate its percentile γ of p0(0.1) j Based on the initial sample u0, event F0 is defined.

[0125] S33: Iteration

[0126] When γ j >γ, iterate

[0127] 1) From the previous level sample u j Select N s = p0N low-value samples, and then use adaptive MCMC to iteratively generate new samples u for each sample j+1 , and updates the value of the MCMC scaling β.

[0128] 2) Recalculate geval j+1 and γ j+1 .

[0129] S34: Estimating the probability of failure

[0130] The failure probability is estimated according to the following formula Finally, the number of samples whose security assessment function value is less than γ in the last sampling is recorded as N f

[0131]

[0132] Ultimately, we average the probabilities obtained from the 30 simulations as our final failure probability estimate.

[0133] S4: Acceleration effect evaluation and optimization

[0134] This step evaluates the algorithm performance from two aspects: the effectiveness and efficiency of the rare event simulation algorithm.

[0135] S41: Algorithm Validity

[0136] The effectiveness of the algorithm is proved by calculating the relative error between the evaluation results of the SuS method combined with the ApMCMC method and the true probability value.

[0137] The fault detection threshold is set to γ ​​= 1, and the true value is approximated by 5 million random Monte Carlo simulation iterations. The SuS algorithm is used to repeatedly estimate the algorithm failure probability 30 times, and each evaluation result is recorded. The results are shown in Figure 5(a) to Figure 5(b) As shown in the figure, the relative errors of the final estimates obtained through the experiment are 0.65% and 2.84%, respectively, both within a small range. This shows that the example algorithm can accurately estimate the probability of rare events, thereby effectively evaluating the failure probability of the intelligent decision-making algorithm in the vehicle-following scenario.

[0138] Figure 5(a) to Figure 5(b) In the figure, the horizontal axis represents the number of repeated experiments; the vertical axis represents the estimated value of the failure probability of this time; Pf-episode represents the estimated value of the failure probability in each repeated experiment; Pf-mean represents the average value of the estimated failure probability of thirty repeated experiments; Pf-real represents the true value of the failure probability of the algorithm.

[0139] S41: Algorithm efficiency

[0140] Under the same test sample, the effectiveness of the algorithm can be proved by comparing the example algorithm with the traditional probability evaluation method, the naive Monte Carlo (MC) method. This example compares SuS+ApMCMC with the MC method in terms of estimation accuracy, estimation variance, and confidence interval. The comparison of estimation error and confidence interval is shown in Figure 6. The experiment proves that the proposed method has several to dozens of times advantages over the traditional Monte Carlo method in the above indicators.

[0141] In addition, by comparing the confidence interval and variance ratio of MC, SuS, SuS+ApMCMC and another classic rare event simulation algorithm Sequential Importance Simulation (SiS) at different thresholds, as shown in Figure 7, it can be found that the SiS algorithm underestimates the failure probability under low thresholds, and ApMCMC+SuS shows a significantly smaller variance among the other three algorithms, indicating that it has excellent evaluation performance, proving the efficiency of the algorithm.

[0142] Figure 6(a) to Figure 7(b) In , the horizontal axis represents the value of the failure probability γ; the vertical axis of Figure 6(a) represents the estimated error value under each failure threshold; Figure 6(b) to Figure 7(a)The ordinate represents the estimated value of failure probability at each failure threshold; the ordinate of Figure 7(b) represents the estimated variance ratio between the Monte Carlo algorithm and the experimental algorithm.

[0143] In FIG6( a ), “MC for DDPG” represents the experimental results of applying the MC algorithm to the DDPG algorithm; and “SuS for DDPG” represents the experimental results of applying the SuS algorithm to the DDPG algorithm.

[0144] In Figure 6(b), “99%Confidence Interval for MC” represents the 99% confidence interval estimated by the MC algorithm; “99%Confidence Interval for SuS” represents the 99% confidence interval estimated by the SuS algorithm; “Probalility Real” represents the true probability value; γ test It represents the horizontal axis.

[0145] In Figure 7(a), “SuS with Adaptive MCMC” represents the estimation result using the ApMCMC+SuS algorithm; “SuS with MCMC” represents the estimation result using the MCMC+SuS algorithm; “SiS withAdaptive MCMC” represents the estimation result using the ApMCMC+SiS algorithm; “Probability Real forDDPG” represents the true probability value of the DDPG algorithm; γ test It represents the horizontal axis.

[0146] In Figure 7(b), “MC / SuS with Adaptive MCMC” represents the ratio of the estimated variance between the MC algorithm and the ApMCMC+SuS algorithm; “MC / SuS with MCMC” represents the ratio of the estimated variance between the MC algorithm and the MCMC+SuS algorithm; and “SiS withAdaptive MCMC” represents the ratio of the estimated variance between the MC algorithm and the ApMCMC+SiS algorithm.

[0147] So far, we have completed the accelerated evaluation of the credibility evaluation of the decision-making algorithm for human-vehicle-road collaborative unmanned driving. Specifically, we took the classic deep reinforcement learning DDPG (Deep Deterministic Policy Gradient) and A3C (Asynchronous Advantage Actor Critic) as examples. The evaluation results are as follows: Figure 8 shown.

[0148] Figure 8In the figure, “99%Confidence Interval for DDPG MC” indicates the 99% confidence interval of the failure probability of the DDPG algorithm using the MC algorithm;

[0149] “99%Confidence Interval for DDPG SuS” means the 99% confidence interval of the failure probability of the DDPG algorithm using the SuS algorithm; “99%Confidence Interval for A3C MC” means the 99% confidence interval of the failure probability of the A3C algorithm using the MC algorithm; “99%Confidence Interval for A3C SuS” means the 99% confidence interval of the failure probability of the A3C algorithm using the SuS algorithm; “Probability Real for A3C” means the real failure probability of the A3C algorithm; “Probability Real for DDPG” means the real failure probability of the DDPG algorithm.

Claims

1. A method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving, characterized in that: The following steps are involved: 1) Construction of typical test scenarios for autonomous driving: 1.1) Construction of a black box system test platform for unmanned driving: Based on real data sets, typical application scenarios for unmanned driving are built. Based on the selected typical application scenarios for unmanned driving, the scenario elements are analyzed. Based on the abstract scenarios, the scenario parameters are further mined and logical scenarios are constructed. Based on real data sets, data preprocessing is performed to extract the driving behavior of vehicles in specific application scenarios as the data basis; 1.2) Construction of parameter distribution of typical operation scenarios: Based on the extracted vehicle driving data and scenario parameter division, the basic distribution of the parameters is fitted, and specific test scenarios are generated through distribution sampling technology as input variables for subsequent tests; 2) Single test scenario evaluation: 2.2) Autonomous driving black box algorithm training: Use deep reinforcement learning algorithms to replace the unmanned driving black box algorithms as the evaluation object, and design reward functions based on the operating requirements of the unmanned driving system in specific application scenarios to make the algorithm converge; 2.2) Failure assessment function design: Trustworthy evaluation of the performance of the unmanned driving black box algorithm in the human-vehicle-road collaborative environment, introducing a trustworthy evaluation index system; comprehensively evaluating the trustworthy evaluation indicators, further designing a function to measure the algorithm performance in a specific parameter scenario, and after the simulation is completed, calculating the function value based on the simulation output data, and comparing it with the preset failure judgment threshold to determine the performance score of the algorithm in this scenario and whether it fails; 3) System acceleration test evaluation: 3.1) Subset simulation combined with adaptive Markov chain Monte Carlo method: Based on the probability estimation method, a set simulation combined with an adaptive Markov chain Monte Carlo method is proposed to accelerate the evaluation of the failure probability of the unmanned driving black box algorithm in a human-vehicle-road collaborative environment in a simulation environment; 3.2) Failure cause analysis: By observing the change size and direction of each dimension of sample parameters during sample iteration, the failure cause analysis of the unmanned driving black box algorithm can be realized; 4) Acceleration effect evaluation and optimization: 4.1) Evaluation of algorithm effectiveness: From the perspective of algorithm effectiveness, evaluate and tune the acceleration effect of the credibility acceleration evaluation algorithm, including calculating the relative error between the evaluation result of the selected method and the true probability value for evaluation; 4.2) Evaluation of algorithm efficiency: The acceleration effect evaluation and tuning of the credibility acceleration evaluation algorithm are carried out from the perspective of algorithm efficiency, including calculating the differences between the selected method and the traditional Monte Carlo method in estimation accuracy, estimation variance, confidence interval, and calculation time, as well as the differences between the selected method and other rare event simulation algorithms in the above indicators.

2. The method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving according to claim 1, characterized in that: The construction of the unmanned driving black box system test platform includes the following steps: 2.1) Analyze scene composition: Analyze scene elements based on the selected typical application scenarios of unmanned driving; 2.2) Mining scene parameters: Further mining scene parameters based on abstract scenes to construct logical scenes; 2.3) Extract vehicle-specific behaviors: Based on the real vehicle operation dataset, extract the vehicle’s driving behavior in specific application scenarios.

3. The method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving according to claim 1, characterized in that: The unmanned driving black box algorithm training includes the following steps: 3.1) Training data selection: Extract vehicle behaviors based on real data sets as training data set input; 3.2) Reward function design: Design the reward function based on safety, efficiency, and comfort indicators; 3.3) Score indicator convergence: Modify the hyperparameters and train the model until the score converges.

4. The method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving according to claim 1, characterized in that: The subset simulation combined with the adaptive Markov chain Monte Carlo method comprises the following steps: 4.1) Initialization settings: including initialization parameters and designing distribution transformation functions; 4.2) Generate initial samples: Generate initial samples using random Monte Carlo method; 4.3) Iterative update: Iteratively update samples based on pre-defined evaluation criteria; 4.4) Estimate the failure probability: Calculate the failure probability according to the formula.

5. The method for accelerating the credibility evaluation of human-vehicle-road collaborative unmanned driving according to claim 1, characterized in that: The algorithm efficiency evaluation includes the following steps: 5.1) Comparison with the traditional Monte Carlo method: comparison in terms of estimation accuracy, estimation variance, confidence interval, calculation time, etc.; 5.2) Comparison with other rare event simulation algorithms: Compare with SuS algorithm and other rare event simulation algorithms such as importance sampling and sequence importance sampling methods in various dimensions.

Images