An explainable decision-making method for autonomous driving based on recommendation algorithm

Abstract

The present invention discloses an explainable decision-making method for autonomous driving based on a recommendation algorithm, which mainly relates to the field of recommendation algorithms and autonomous driving; the method comprises the following steps: S1, extracting alternative indicators and driving behavior characteristics of evaluation indicators after preprocessing traffic data obtained by sensors, and clustering to obtain labels; S2, training an XGBoost model to analyze feature importance, screening key driving behavior characteristics of each evaluation indicator through a recursive feature elimination test, and retraining the model; S3, for key driving behavior characteristics of high-risk traffic data, recalling multiple groups of key feature parameter combinations according to historical data, and clustering to form groups with different contribution levels; S4, performing multi-objective constraint operations research based on the recommended key feature parameter combinations and their performance prediction values; compared with existing deep learning algorithms, the method has higher explainability and is conducive to the popularization and use of autonomous driving technology.

Description

Technical Field

[0001] The present invention relates to the field of recommendation algorithms and autonomous driving, and specifically to an explainable decision-making method for autonomous driving based on a recommendation algorithm. Background Art

[0002] With the development of autonomous driving technology, how to significantly improve traffic safety and efficiency has become an urgent problem to be solved. In autonomous driving technology, autonomous driving decision-making is the core link to ensure that vehicles can travel safely and efficiently. However, autonomous driving decisions involve multiple conflicting goals, such as risk, efficiency, and passenger comfort. Existing single-objective decision-making algorithms are difficult to dynamically adapt to different driving scenarios and environmental conditions. Decision-making models that comprehensively consider multiple goals such as risk, comfort, and efficiency need to balance between multiple goals, making decision-making inefficient. In addition, explainability is the basis for building user trust, but due to the "black box" characteristics of deep learning networks, existing autonomous driving decision-making models generally have insufficient explainability. The lack of transparency in the model makes it difficult to understand the decision-making process, which in turn produces unreliable and unpredictable decision results. Summary of the invention

[0003] The purpose of the present invention is to solve the problems existing in the prior art and to provide an explainable decision-making method for autonomous driving based on a recommendation algorithm. Compared with the existing deep learning algorithm, the method has higher explainability and is conducive to the promotion and use of autonomous driving technology.

[0004] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0005] An explainable decision-making method for autonomous driving based on a recommendation algorithm, comprising the steps of:

[0006] S1. After preprocessing the traffic data obtained by the sensor, the alternative indicators of the evaluation index and the driving behavior characteristics are extracted, and the labels are obtained by clustering;

[0007] S2. Train the XGBoost model to analyze feature importance, select key driving behavior features of each evaluation index through recursive feature elimination test and retrain the model;

[0008] S3. For the key driving behavior characteristics of high-risk traffic data, multiple groups of key feature parameter combinations are recalled based on historical data, and clustered into groups with different contribution levels;

[0009] S4. Based on the recommended key feature parameter combinations and their performance prediction values, multi-objective constraint operations research is performed.

[0010] Preferably, in step S1, the step of preprocessing the traffic data obtained by the sensor includes data type conversion, data cleaning, and time period data splicing.

[0011] Preferably, in step S1, the evaluation indicators of driving data are risk, driving efficiency, and comfort, and alternative indicators and driving behavior characteristics of each indicator are extracted, wherein the alternative indicator of risk is defined as collision time, time-comprehensive collision time, collision avoidance deceleration, and potential collision index; the alternative indicator of driving efficiency is defined as headway; the alternative indicator of comfort is defined as jerk and acceleration;

[0012] The collision time is calculated as follows:

[0013]

[0014] Among them, x i (t) and v i (t) is the position and speed of the current vehicle (i) at time stamp t, L i-1 is the length of the preceding vehicle (i-1);

[0015] The calculation formula of time integrated collision time is as follows:

[0016]

[0017] The calculation formula for collision avoidance deceleration is as follows:

[0018]

[0019] The calculation formula of the potential collision index is as follows:

[0020]

[0021] Among them, MADR is specific to a given set of traffic and environmental attributes (α1, α2, …, α n ), which depends on factors including road conditions, vehicle weight and braking system.

[0022] Preferably, in step S1, based on the extracted risk, driving efficiency and comfort alternative indicators, the traffic data obtained by the sensor is clustered by a fuzzy C-means clustering algorithm to obtain risk, driving efficiency and comfort labels.

[0023] Preferably, in step S2, based on the labels of each evaluation indicator obtained by clustering in step S1, an XGBoost model of each evaluation indicator is trained, and features with higher feature importance of each evaluation indicator are obtained.

[0024] Preferably, in step S2, features are eliminated one by one through a recursive feature elimination test and the evaluation model is retrained to determine the difference in prediction performance of different feature inputs, and the key features at the inflection point are selected by comprehensively considering the number of features and the prediction performance.

[0025] Preferably, in step S2, all key driving behavior features are used to retrain the XGBoost models for risk, driving efficiency, and comfort scores, respectively.

[0026] Preferably, in step S3, based on the risk labels evaluated in step S1, the key driving behavior features of the high-risk traffic data are adjusted, and the data distribution of the key driving behavior features in the low-risk traffic data is statistically analyzed to limit the data search range; the data distribution of the key driving behavior features in the low-risk traffic data is used as the recall basis, and the XGBoost model trained in step S2 is used to predict the performance indicators corresponding to various recall parameter combinations, and the key feature parameter combinations are screened and finely sorted according to the predicted performance indicators.

[0027] Preferably, in step S3, the key feature data combinations with good performance indicators are clustered by using the KMeans clustering algorithm to form groups with different contribution levels.

[0028] Preferably, in step S4, different weight coefficients are set for the three evaluation indicators, the objective function and constraints are defined, and the optimization algorithm is run within the framework of the constraints to search for the Pareto optimal solution;

[0029] The objective function is calculated as follows:

[0030]

[0031] Among them, w1, w2, and w3 are the weights of risk index, driving efficiency index, and comfort index respectively;

[0032]

[0033] Among them, x i is a binary decision variable, indicating whether to select the i-th group of parameters.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] The present invention can effectively screen out key features of risk, driving efficiency and comfort through recursive feature elimination tests, and train the model again based on these key features. Compared with existing deep learning algorithms, the present invention has higher interpretability and is more comprehensive than existing single-objective optimization autonomous driving decision-making methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the present invention;

[0037] Figure 2 is a comparison chart of the predicted scores of risk features;

[0038] Figure 3It is a comparison chart of predicted scores of driving efficiency;

[0039] Figure 4 This is a comparison chart of the predicted scores of comfort;

[0040] Figure 5 It is a vehicle driving data evaluation form. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0042] Example: As attached Figure 1 As shown, the present invention is an explainable decision-making method for autonomous driving based on a recommendation algorithm, comprising:

[0043] Step 1: After preprocessing the traffic data obtained by the sensor, extract alternative indicators of the evaluation index and driving behavior characteristics, and cluster them to obtain labels.

[0044] Furthermore, the preprocessing steps of the traffic data obtained by the sensor in step 1 include data type conversion, data cleaning, and time-divided data splicing.

[0045] Furthermore, the evaluation indicators of driving data in step 1 are risk, driving efficiency, and comfort, and the alternative indicators and driving behavior characteristics of each indicator are extracted. The alternative indicators of risk are defined as time to collision (TTC), time-integrated collision time (TIT), deceleration to avoid collision (DRAC), and potential collision index (CPI), the alternative indicators of driving efficiency are defined as headway (HdwyT), and the alternative indicators of comfort are defined as jerk and acceleration (Acc).

[0046] The time to collision (TTC) calculation formula is as follows:

[0047]

[0048] Among them, x i (t) and v i (t) is the position and speed of the current vehicle (i) at time stamp t, L i-1 is the length of the preceding vehicle (i-1).

[0049] The time-integrated collision time (TIT) is calculated as follows:

[0050]

[0051] The calculation formula for collision avoidance deceleration (DRAC) is as follows:

[0052]

[0053] The calculation formula of the collision potential index (CPI) is as follows:

[0054]

[0055] Among them, MADR is specific to a given set of traffic and environmental attributes (α1, α2, …, α n ) and depends on factors such as road surface conditions and vehicle weight and braking system (Cunto and Saccomanno, 2008).

[0056] Furthermore, in step 1, based on the extracted risk, driving efficiency, and comfort alternative indicators, the traffic data obtained by the sensor is clustered by the fuzzy C-means clustering algorithm to obtain risk, driving efficiency, and comfort labels.

[0057] Step 2: Train the XGBoost model to analyze feature importance, screen the key driving behavior features of each evaluation indicator through recursive feature elimination experiments, and retrain the model.

[0058] Furthermore, in step 2, based on the labels of each evaluation indicator obtained by clustering in step 1, the XGBoost model of each evaluation indicator is trained, and the features with higher feature importance of each evaluation indicator are obtained.

[0059] Furthermore, in step 2, through recursive feature elimination experiments, features are eliminated one by one and the evaluation model is retrained to determine the difference in prediction performance of different feature inputs, and the key features at the inflection point are selected by comprehensively considering the number of features and the prediction performance.

[0060] Furthermore, all key driving behavior features are used in step 2 to retrain the XGBoost models for risk, driving efficiency, and comfort scores respectively.

[0061] Step 3: For the key driving behavior features of high-risk traffic data, recall multiple groups of key feature parameter combinations based on historical data and cluster them into groups with different contribution levels.

[0062] Furthermore, step 3 adjusts the key driving behavior features of the high-risk traffic data based on the risk labels evaluated in step 1, and counts the data distribution of the key driving behavior features in the low-risk traffic data to limit the data search scope.

[0063] Furthermore, in step 3, the data distribution of key driving behavior characteristics in low-risk traffic data is used as the basis for recall.

[0064] The XGBoost model trained in step 2 is used to predict the performance indicators corresponding to various recall parameter combinations, and the key feature parameter combinations are screened and finely sorted according to the predicted performance indicators.

[0065] Furthermore, in step 3, the KMeans clustering algorithm is used to cluster the key feature data combinations with good performance indicators to form groups with different contribution levels.

[0066] Step 4: Based on the recommended key feature parameter combinations and their performance prediction values, multi-objective constraint operations research is performed.

[0067] Furthermore, in step 4, different weight coefficients are set for the three evaluation indicators, the objective function and constraints are defined, and the optimization algorithm is run within the framework of the constraints to search for the Pareto optimal solution.

[0068] The objective function is calculated as follows:

[0069]

[0070] Among them, w1, w2, and w3 are the weights of the risk index, driving efficiency index, and comfort index respectively.

[0071]

[0072] Among them, x i is a binary decision variable, indicating whether to select the i-th group of parameters.

[0073] The key features of risk, driving efficiency and comfort can be effectively screened out through recursive feature elimination test. The model is trained again based on these key features. Figure 2-4 It can be seen that the present invention is more interpretable than the existing deep learning algorithms.

[0074] Attached Figure 5 In the above analysis, the driving data of vehicles with higher risk levels are selected for evaluation. RiskLabel, EfficiencyLabel, and ComfortLabel are the evaluation indexes before optimization, and RiskLabelOpt, EfficiencyLabelOpt, and ComfortLabelOpt are the evaluation indexes after optimization. It can be clearly seen from the first 10 groups of vehicle data that the risk level is significantly reduced and the comfort index is improved when driving efficiency allows, which is more comprehensive than the existing single-objective optimization autonomous driving decision-making method.

Claims

1. An explainable decision-making method for autonomous driving based on a recommendation algorithm, characterized in that: Includes steps: S1. After preprocessing the traffic data obtained by the sensor, the alternative indicators of the evaluation index and the driving behavior characteristics are extracted, and the labels are obtained by clustering; The evaluation indicators of driving data are risk, driving efficiency, and comfort, and the alternative indicators and driving behavior characteristics of each indicator are extracted. Among them, the alternative indicators of risk are defined as collision time, time-comprehensive collision time, collision avoidance deceleration, and potential collision index; the alternative indicator of driving efficiency is defined as headway; the alternative indicator of comfort is defined as jerk and acceleration; The collision time is calculated as follows: Among them, x i (t) and v i (t) is the position and speed of the current vehicle (i) at time stamp t, L i-1 is the length of the preceding vehicle (i-1); The calculation formula of time integrated collision time is as follows: The calculation formula for collision avoidance deceleration is as follows: The calculation formula of the potential collision index is as follows: Among them, MADR is specific to a given set of traffic and environmental attributes (α1, α2, …, α n ), which depends on factors including road conditions, vehicle weight and braking system; S2. Train the XGBoost model to analyze feature importance, select key driving behavior features of each evaluation index through recursive feature elimination test and retrain the model; Based on the labels of each evaluation indicator obtained by clustering in step S1, train the XGBoost model of each evaluation indicator and obtain the features with higher feature importance for each evaluation indicator; Through recursive feature elimination experiments, features are eliminated one by one and the evaluation model is retrained to determine the difference in prediction performance of different feature inputs. The key features at the inflection point are selected by comprehensively considering the number of features and prediction performance. Retrain the XGBoost models for risk, driving efficiency, and comfort scores using all key driving behavior features; S3. For the key driving behavior characteristics of high-risk traffic data, multiple groups of key feature parameter combinations are recalled based on historical data, and clustered into groups with different contribution levels; Based on the risk labels evaluated in step S1, the key driving behavior features of the high-risk traffic data are adjusted, and the data distribution of the key driving behavior features in the low-risk traffic data is counted to limit the data search scope; the data distribution of the key driving behavior features in the low-risk traffic data is used as the recall basis, and the XGBoost model trained in step S2 is used to predict the performance indicators corresponding to various recall parameter combinations, and the key feature parameter combinations are screened and finely sorted according to the predicted performance indicators; The KMeans clustering algorithm is used to cluster key feature data combinations with good performance indicators to form groups with different contribution levels; S4. Based on the recommended key characteristic parameter combination and its performance prediction value, multi-objective constraint operations research is performed; Set different weight coefficients for the three evaluation indicators, define the objective function and constraints, and run the optimization algorithm within the framework of the constraints to search for the Pareto optimal solution; The objective function is calculated as follows: Among them, w1, w2, and w3 are the weights of risk index, driving efficiency index, and comfort index respectively; Among them, x i is a binary decision variable, indicating whether to select the i-th group of parameters.

2. The autonomous driving explainable decision-making method based on the recommendation algorithm according to claim 1, characterized in that: In step S1, the steps of preprocessing the traffic data obtained by the sensor include data type conversion, data cleaning, and time-divided data splicing.

3. The autonomous driving explainable decision-making method based on the recommendation algorithm according to claim 1, characterized in that: In step S1, based on the extracted risk, driving efficiency, and comfort alternative indicators, the traffic data obtained by the sensor is clustered by the fuzzy C-means clustering algorithm to obtain risk, driving efficiency, and comfort labels.

Images