A human-machine decision-making logic online optimization method for high-level automatic driving

Abstract

The application discloses a kind of human-machine decision logic online optimization methods for high degree of automatic driving, its method is: first step, human-machine hybrid enhanced decision database online optimization;Second step, human-machine hybrid enhanced decision model online optimization;Third step, human-machine hybrid enhanced decision prediction model online optimization;Beneficial effect: improve the credibility of human-machine hybrid enhanced decision result and the safety of human-machine co-driving vehicle travel.Maximize to human-machine hybrid enhanced decision database is improved, give full play to the advantage of vehicle network information sharing.Maximize to human-machine hybrid enhanced decision model is improved, give full play to the advantage of vehicle network information sharing.Greatly improve the prediction effect to human-machine co-driving vehicle future state, guarantee the safety of human-machine co-driving vehicle travel.

Description

Technical Field

[0001] This invention relates to an online optimization method for human-machine decision-making logic, specifically an online optimization method for human-machine decision-making logic for highly automated driving. Background Technology

[0002] Currently, due to the technological limitations and social dilemmas of fully autonomous driving, coupled with low public acceptance, human-machine co-driving will become a long-term form of intelligent driving in the future. In-depth analysis of the human-machine consistency and collaboration mechanism is the theoretical foundation of human-machine co-driving. Influenced by the "human driving" attribute and the "machine driving" decision-making logic, research on safe and reliable human-machine co-driving based on highly consistent human-machine decision-making is one of the core technologies of future intelligent transportation systems. Currently, the decision-making layer of human-machine co-driving systems mainly relies on data training to realize the hybrid decision-making process and human-machine driving right arbitration logic. Data training algorithms based on second- and third-generation artificial intelligence theories generally suffer from the drawbacks of being unexplainable and uninterpretable. Furthermore, online decision-making failures pose safety hazards, and the weak self-learning ability of decision-making algorithms limits the vehicle's adaptability to different scenarios. These system defects collectively challenge the reliability of the human-machine hybrid decision-making process and its output results. Therefore, it is necessary to establish an online optimization method for human-machine decision-making logic for highly automated driving, optimizing the human-machine hybrid enhanced decision database, human-machine hybrid enhanced decision model, and human-machine hybrid enhanced decision prediction model online to ensure the safety and reliability of the decision results.

[0003] Online optimization of human-machine decision-making logic for highly automated driving has several characteristics. First, it demands high real-time performance. Real-world driving environments are highly complex and dynamically changing. When driving situations or scenarios not present in the human-machine hybrid augmented decision database arise, the human-machine hybrid augmented decision model must be able to quickly output safe and reasonable decisions within a very short time. Therefore, the online optimization model for human-machine decision-making logic needs strong real-time performance to ensure the safety of human-machine co-driving vehicles. Second, it has high model complexity. Online optimization of human-machine decision-making logic needs to consider multiple factors and optimize multiple modules related to human-machine hybrid augmented decision-making, such as the human-machine hybrid augmented decision database, the human-machine hybrid augmented decision model, and the human-machine hybrid augmented decision prediction model. Therefore, the online optimization model for human-machine decision-making logic has high complexity. Finally, it places high demands on the computing power, storage capacity, and network transmission capabilities of the equipment. Due to the high real-time requirements and model complexity of online optimization of human-machine decision-making logic, the computing power, storage capacity, and network transmission capabilities of the relevant equipment must be sufficiently powerful to meet the requirements of online optimization.

[0004] Currently, domestic and international teams have explored relevant methods for model optimization, but existing technologies still have certain shortcomings. Firstly, there are few technologies for optimizing human-machine decision-making logic in the field of highly automated driving. Most research focuses on establishing hybrid human-machine decision-making models, lacking research on the reliability evaluation and optimization methods of these models. Secondly, there are few technologies for online optimization of human-machine decision-making logic. Traffic systems differ from other systems, exhibiting highly nonlinear dynamic characteristics and extremely stringent safety requirements, necessitating online optimization models with high real-time performance. Finally, optimization models for hybrid human-machine augmented decision-making are relatively simple and lack a systematic approach. While current hybrid human-machine augmented decision-making models have incorporated machine learning algorithms from artificial intelligence fields such as deep learning and reinforcement learning, the optimization models still remain at the level of adjusting decision-making methods and model parameters based on test results, failing to meet the requirements of online and real-time performance, and lacking a complete online optimization system for human-machine decision-making logic.

[0005] Chinese patent CN201911358599.5 discloses a dynamic human-machine co-driving driving right allocation method based on real-time driver risk response. This method involves collecting environmental and vehicle information to obtain the comprehensive risk intensity of surrounding vehicles on the vehicle, thereby judging the driver's driving state in real time and calculating the driving weight allocation factor. It belongs to a switching-type human-machine co-driving decision model. Chinese patent 201910154814.3 discloses a driving right allocation method in a human-machine co-driving lane-keeping system. It uses fuzzy control to determine the co-driving coefficient, achieving continuous change of the co-driving coefficient to prevent sudden changes in control right and ensure the safety and comfort of the lane-keeping system. It belongs to a shared-type human-machine co-driving decision model. Both patents establish human-machine hybrid decision-making models, but they do not perform online optimization of the designed models, and therefore cannot fully guarantee the reliability of the decision results and driving safety. Chinese Patent 202110090864.7 discloses a human-machine co-driving test method based on digital twin virtual-real interaction technology. This method tests and optimizes the vehicle's human-machine interaction, co-driving performance, and passenger comfort through the interaction between virtual simulation scenarios, virtual sensors (radar, cameras, etc.), and the actual vehicle's autonomous driving controller. While this patent verifies and optimizes the human-machine hybrid decision-making model from a simulation experiment perspective, it does not improve the mechanism of the decision-making logic, thus failing to meet real-time requirements and lacking online optimization capabilities. Summary of the Invention

[0006] The purpose of this invention is to address the issue that the technology for optimizing human-machine decision logic in the field of highly automated driving cannot meet the requirements of online and real-time performance, and to provide an online optimization method for human-machine decision logic in highly automated driving.

[0007] The present invention provides an online optimization method for human-machine decision-making logic in highly automated driving, the method comprising the following steps:

[0008] The first step is to optimize the online decision-making database using a human-machine hybrid approach. The specific steps are as follows:

[0009] Step 1: Update the human-machine hybrid augmented decision database by sensing new driving situations or scenarios through sensors. Process and fuse the sensor information, convert it into relevant knowledge content by the knowledge acquisition module, and then determine whether the knowledge content belongs to a new knowledge type by the knowledge judgment module. If so, update the human-machine hybrid augmented decision database for online optimization.

[0010] Step 2: Update the human-machine hybrid augmented decision database by downloading from the cloud. The updated human-machine hybrid augmented decision database of each human-machine co-driving vehicle connected to the cloud is uploaded to the cloud in real time to enrich the content of the cloud database. Then, the knowledge judgment module updates the content of the cloud database to all human-machine co-driving vehicles.

[0011] Step 2: Online optimization of the human-machine hybrid augmented decision-making model;

[0012] Step 1: Optimize the human-machine hybrid augmented decision-making model online by consulting with drivers. Construct new knowledge reasoning between the updated knowledge in the human-machine hybrid augmented decision-making database and the decision results. Then, use this new knowledge reasoning to optimize the human-machine hybrid augmented decision-making model online, achieving incremental learning of the human-machine hybrid augmented decision-making database. Step 1 is completed in two stages, as follows:

[0013] Step 1: Evaluating the Model's Decision-Making Capability: After receiving sensory input, the human-machine hybrid decision-making model will pass through different reasoning modules in the human-machine hybrid augmented decision-making model. Each reasoning module will determine in turn whether it has the ability to make a decision output for the sensory input. When all reasoning modules in the human-machine hybrid augmented decision-making model are unable to provide knowledge reasoning for making a behavioral decision based on the sensory input, it indicates that the current human-machine hybrid augmented decision-making model has poor decision-making capability and cannot cope with the current sensory input situation. In this case, the model will turn to the human driver for consultation, and the human driver will provide specific instructions.

[0014] Step 2: Online Model Optimization: After the human driver gives the driving operation, the human-machine hybrid augmented decision model records the human driver's driving operation and establishes decision knowledge reasoning between the perceptual input and the corresponding human driver's driving operation by learning the perceptual input and the decision output. This allows for online optimization of the case-based reasoning module, rule-based reasoning module, and neural network-based reasoning module in the decision model.

[0015] Step 2: Optimize the human-machine hybrid augmented decision-making model online via cloud download. This involves uploading the updated human-machine hybrid augmented decision-making model of each human-machine co-driving vehicle connected to the cloud in real time, enriching the cloud decision model library. Then, the model judgment module updates the content of the cloud decision model library to all human-machine co-driving vehicles. By updating the decision model through the cloud, the human-machine co-driving vehicles can optimize their own human-machine hybrid augmented decision-making model by learning the knowledge reasoning between the perception input and decision output of other vehicles.

[0016] Step 3: Online optimization of the human-machine hybrid augmented decision prediction model:

[0017] The parameters of the human-machine hybrid augmented decision prediction model are optimized by means of a genetic algorithm. First, the model parameters to be optimized are determined and an objective function for parameter optimization is established. Then, the optimal solution is searched in the solution space by the genetic algorithm until the population converges. Finally, the optimal individual is decoded to obtain the parameter combination, which is used to replace the parameters in the model to complete the online optimization of the human-machine hybrid augmented decision prediction model.

[0018] The specific process of step one in the first step is as follows:

[0019] 1) The Human-Machine Hybrid Augmented Decision Database is part of the Human-Machine Hybrid Augmented Decision Knowledge Base Management System. It includes a rule base, a meta-knowledge base, a professional vocabulary base, and a multi-level knowledge unit base. The rule base stores various relevant rules and regulations for traffic and driving rules of human-machine co-driving vehicles; the meta-knowledge base stores the priority, category, and macro-level description of knowledge to facilitate knowledge manipulation; the professional vocabulary base stores various professional terms related to human-machine co-driving vehicles; and the multi-level knowledge unit base stores knowledge information at various levels. If the knowledge content in the Human-Machine Hybrid Augmented Decision Database is too small, there may be no solution to the decision or the retrieved decision solution may be difficult to adapt to new driving situations and scenarios. Therefore, the Human-Machine Hybrid Augmented Decision Database needs to have online update and optimization functions to lay the foundation and provide a basis for safe and reliable human-machine hybrid augmented decision-making.

[0020] 2) Updating the database by sensing new driving situations or scenarios through sensors is a kind of database optimization method of single-vehicle self-learning evolution. When new driving situations or scenarios exist in the driving environment of human-machine co-driving vehicles, the decision-making ability of the human-machine hybrid augmented decision system often exceeds its boundaries. Therefore, in order to ensure the safety of human-machine hybrid augmented decision-making, it is necessary to record the new driving situations and scenarios, convert them into new knowledge content and store them in the human-machine hybrid augmented decision database, update and enrich the database content, and provide a basis for human-machine hybrid augmented decision-making.

[0021] 3) Human-machine co-driving vehicles perceive their own position, status, and information about the surrounding environment through various types of sensors. These sensors are categorized into three types: positioning sensors, vision sensors, and radar sensors. Positioning sensors fully utilize information collected by GPS and inertial measurement unit (IMU) sensors. High-precision GPS positioning results are used to control the system drift of the IMU, reducing the accumulation of measurement errors. Simultaneously, encoders measure the vehicle's travel distance, enabling position and absolute pose estimation and outputting the vehicle's position and pose information. Vision sensors acquire image information of the surrounding environment through vision cameras. By analyzing the pixel attributes of the images, they infer the attributes and states of objects in the environment. For example, cameras can identify traffic participants in the surrounding environment, outputting their type and state information; they can identify traffic signs, outputting their type and content information; and they can capture street scenes, compare them with navigation maps, and output the vehicle's precise location information. Radar sensors, such as lidar and millimeter-wave radar, can measure the state of dynamic and static traffic participants around the human-machine co-driving vehicle, outputting their speed, distance, and orientation information.

[0022] 4) After acquiring environmental information around the human-machine co-driving vehicle, various types of sensors process and combine the information collected by multiple sensors. This enables the vehicle's perception layer to sense the surrounding environment with low false alarm rate and high output efficiency, thereby reducing blind spots, improving detection accuracy, and simplifying the dimensions of data output. This provides a concise and reliable data foundation for further functional decisions, ensuring the driving safety of the human-machine co-driving vehicle. Sensor fusion consists of three parts: time synchronization, spatial synchronization, and data processing. Time synchronization filters out the information collected by different sensors within a specified time period and sorts them according to their accuracy using interpolation and extrapolation methods. This achieves unification of different sensors at the same time; spatial synchronization solves the coordinate transformation relationship between different sensors through the calibration of a single sensor and the joint calibration of several sensors. After spatial synchronization is completed, the information of the same object under different sensors is mapped to the same coordinate system, thereby solving the information collected by different sensors into the same coordinate system; data processing handles the parts with large differences, data redundancy and data missing in the synchronized sensor measurements to achieve data compatibility of each sensor. Then, the data of the same object measured by different sensors are processed, and the measurement value with the highest accuracy is selected as the sensor output from several measurements, simplifying the data output dimension;

[0023] 5) The information fused from the sensors is input into the knowledge acquisition module of the human-machine hybrid augmented decision-making knowledge base management system. The knowledge acquisition module can convert the information from the sensor-fused human-machine co-driving vehicle and its surrounding environment into knowledge of driving knowledge professional terms, driving knowledge experience rules, driving knowledge rule knowledge units, and behavioral decision meta-knowledge units, and input it into the knowledge judgment module. The knowledge judgment module can call the knowledge in the rule base, meta-knowledge base, professional terminology base, and multi-level knowledge unit base of the current human-machine hybrid augmented decision database and compare it with the knowledge input in the knowledge acquisition module. When there is a new knowledge type, the knowledge judgment module writes the new knowledge content into the corresponding professional terminology base, experience rule base, rule knowledge unit base, and meta-knowledge unit base of the human-machine hybrid augmented decision database according to its category for storage, so as to realize the updating and optimization of the human-machine hybrid augmented decision database.

[0024] The process of step two in the first step is as follows:

[0025] Human-machine co-driving vehicles are intelligent connected vehicles that connect to the cloud via a network. They upload their human-machine hybrid decision-making knowledge base to the cloud database in real time and download and update their human-machine hybrid augmented knowledge base from the cloud database in real time. When encountering a new driving situation or scenario, each human-machine co-driving vehicle first updates its human-machine hybrid augmented decision-making database through sensor-based database updates. After the update is complete, the vehicle promptly uploads the updated database content to the cloud via the network. Upon receiving the new knowledge content, the cloud inputs it into the knowledge judgment module 1 of the cloud knowledge base management system. The knowledge judgment module 1 can access knowledge from the rule base, meta-knowledge base, professional vocabulary base, and multi-level knowledge unit base in the current cloud database. The knowledge judgment module 1 compares the new knowledge content with the newly input knowledge content. When a new knowledge type exists, the knowledge judgment module 1 writes the new knowledge content into the corresponding professional vocabulary library, experience rule library, rule knowledge unit library, and meta knowledge unit library in the cloud database according to its category for storage, thereby updating and optimizing the cloud database. At the same time, the knowledge judgment module 2 of the cloud database knowledge base management system calls the human-machine hybrid augmented decision database of all human-machine co-driving vehicles connected to the cloud and compares the content of the cloud database with it. When the human-machine hybrid augmented decision database of a certain human-machine co-driving vehicle is missing a certain knowledge in the cloud database, the knowledge is downloaded to the human-machine hybrid augmented decision database of that human-machine co-driving vehicle, thereby updating and optimizing the human-machine hybrid augmented decision database.

[0026] The specific process of step one in step two is as follows:

[0027] The human-machine hybrid augmented decision-making model includes a case-based reasoning module, a rule-based reasoning module, and a neural network-based reasoning module. The sensor fusion-based perceptual input includes information on the vehicle's driving state, the motion state of surrounding traffic participants, traffic signs, road conditions, weather conditions, and lighting conditions. This information is input into the human-machine hybrid augmented decision-making model and first enters the case-based reasoning module. Upon receiving the input, the case-based reasoning module searches the stored driving case library and calculates the maximum similarity between the perceptual input and the stored cases in the library. The calculation formula is as follows:

[0028]

[0029] Among them, f ij ω represents the similarity between the j-th attribute of the i-th case in the case library and the j-th attribute of the input case. j F represents the weight coefficient of the j-th attribute. i This represents the similarity between the input case and the i-th case;

[0030] When the maximum similarity is below a threshold, it indicates that the case-based reasoning module cannot generate an effective decision output for the current perceived input. The perceived input is then passed to the next-level rule-based reasoning module. Upon receiving the input, the rule-based reasoning module compares it with knowledge in the driving rule base. If an unknown condition exists, it indicates that the rule-based reasoning module cannot generate an effective decision output for the current perceived input. The perceived input is then passed to the next-level neural network-based reasoning module. Upon receiving the input, the neural network-based reasoning module determines whether the perceived input is within the range of values ​​for the neural network's input variables. If the perceived input exceeds the range of values ​​for the neural network's input variables, it indicates that the neural network-based reasoning module cannot generate an effective decision output for the current perceived input. Finally, it consults the human driver for specific instructions.

[0031] The specific process of step two in step one of the second step is as follows:

[0032] 1) The case-based reasoning module performs online optimization by storing perceptual inputs and corresponding human driver operations as new driving cases in the driving case library;

[0033] 2) The rule-based reasoning module performs online optimization by storing the perceptual input and the corresponding human driver's driving operations as new driving rules and driving experience into the driving rule base and driving experience base.

[0034] 3) The neural network-based reasoning module uses a deep learning-based neural network for knowledge reasoning. It inputs the information fused from the sensors of the human-machine co-driving vehicle into the neural network model. Through the model's calculations, it provides accurate decision information. After obtaining new knowledge reasoning between the perceptual input and the corresponding driving operations of the human driver, the neural network-based reasoning module adjusts the brain-like neuron structure, internal model parameter dimensions, and organizational form of the model through evolutionary algorithms, derives the model structure, and completes online optimization.

[0035] The basic computational unit of a neural network is the neuron, and the formula for calculating the output is as follows:

[0036] C=f(XW+b) (2)

[0037] Where X is the input signal of each layer, W is the weight corresponding to the input signal, b is the bias, which can enhance the expressive ability of the neuron, f is the activation function, and C is the output signal of each layer.

[0038] 4) Train a fully connected neural network, continuously adjusting the network's weights and biases under the stimulation of perceived input, so that the network's output continuously approaches the desired decision result. In the training of the neural network, a non-linear thresholded activation function, the Sigmoid function, is chosen, and its calculation formula is as follows:

[0039]

[0040] The formula for calculating the output of the nth hidden layer in a neural network is:

[0041]

[0042] Where x m Let ω be the input vector. nm Let b be the weights corresponding to nodes from the m-th layer to the n-th layer, where N is the number of nodes in the n-th layer. n For the bias of the nth layer, Hide n It is the output of the nth layer;

[0043] The formula for calculating the output layer is:

[0044]

[0045] Where C is the output signal of the output layer, ω cn Here, b represents the weights of the nth hidden layer, P is the total number of hidden layers, and b is the weights of the nth hidden layer. c For the bias of the output layer, f c The activation function for the output layer;

[0046] The residual E is obtained by calculating the mean square error between the output C and the actual label. The calculation formula is as follows:

[0047]

[0048] The weights from the output layer to the hidden layer are updated using gradient descent. The impulse of the weights and biases is calculated using the chain rule, and then both are updated separately. The formulas for calculating the impulse of the weights and biases are as follows:

[0049]

[0050]

[0051] in and Here, η represents the impulse of the weights and biases, respectively; η is the learning rate; and O is the intermediate function.

[0052] 5) Finally, update all weights and biases between the hidden layer and the input layer layer by layer. The formula for calculating the impulse of the weights and biases is as follows:

[0053]

[0054]

[0055] in and These are the impulses for the weights and biases, respectively.

[0056] 6) Evolutionary algorithms encode various parameters of a neural network into chromosomes in the evolutionary algorithm, and then use the evolutionary algorithm to build the network and optimize the parameters. This evolutionary neural network uses the weighted topology evolution method, which automatically evolves a neural network structure that meets the requirements based on the perceived input and the human driver's decision output, including network weights, network structure and activation function.

[0057] 7) The weighted topology evolution method uses an evolutionary neural network based on enhanced topology. First, a complete neural network model including network topology and connection weights is constructed and encoded into node genes and weight genes.

[0058] 8) Perform mutation and crossover operations, and introduce innovation IDs to record the history of gene crossover and mutation. Mutation operations include four parts: adding a node, adding connections between nodes, adjusting weights, and changing the response of the activation function. Crossover operations follow the principle of "matching gene points are used directly in child genes, disconnected gene points or excess genes are obtained from parent genes with high fitness, and two parent genes with the same fitness are randomly inherited."

[0059] 9) Population dissimilarity distance is defined using disjointed gene points, matched gene points, and excess gene points. The more mismatches there are between two genes, the greater the structural difference between them. The calculation formula is as follows:

[0060]

[0061] Where d represents the difference distance, N represents the number of excess gene points in the two genes, M represents the number of disjoint gene points in the two genes, a1, a2 and a3 are coefficients, P is the average of the weight differences of the matching gene points, and L is the length of the longer gene of the two genes. According to this formula, networks with similar structures and weights are assigned to the same group, and intra-group competition occurs during evolution. In this way, populations with new structures do not immediately perish and have sufficient time to adapt to the environment.

[0062] 10) To avoid a situation where some genes in a group have excessively high fitness to the environment, masking the fitness of other genes with low fitness and reducing the diversity of the entire species, explicit fitness is defined. This means that genes belonging to the same group of species share the same fitness to the environment. The formula for calculating explicit fitness is as follows:

[0063]

[0064] Where f i denoted by gene fitness, δ(i,j) represents the difference between two individuals in the group. When δ(i,j) is less than the difference value, sh(δ(i,j)) is 1, and when δ(i,j) is greater than the difference value, sh(δ(i,j)) is 0. n is the total number of individuals in the group.

[0065] 11) By sharing fitness, each gene is constrained. When there are too many members in a group, its fitness evaluation will be reduced. When a new gene is generated, it will be protected. In this way, if the environmental fitness of an old gene does not improve after multiple generations of evolution, it will be gradually eliminated.

[0066] 12) The evolutionary algorithm trains the deep learning-based neural network in the neural network-based reasoning module based on the new knowledge reasoning formed between the perceptual input and the human driver's decision output. It adjusts the internal brain-like neuron structure, internal model parameter dimensions and organization form to derive the model structure, realize incremental learning of the human-machine hybrid enhanced decision database, and complete the online optimization of the neural network-based reasoning module.

[0067] The specific process of step two in the second step is as follows:

[0068] Human-machine co-driving vehicles are intelligent connected vehicles that connect to the cloud via a network. The vehicle's human-machine hybrid augmented decision-making model is uploaded to the cloud decision model library in real time, and updated in real time from the cloud decision model library. Each human-machine co-driving vehicle performs knowledge reasoning between new perceptual inputs and decision outputs, completing online optimization of case-based reasoning, rule-based reasoning, and neural network-based reasoning modules. The newly updated human-machine hybrid augmented decision-making model is then promptly uploaded to the cloud via the network. Upon receiving the input human-machine hybrid augmented decision-making model, the cloud inputs it into the model judgment module 1 of the cloud decision model library management system. The model judgment module 1 compares the input decision model with the corresponding modules in the current cloud decision model library's case-based, rule-based, and neural network-based reasoning modules. When new driving case libraries, driving rule libraries, or neural networks exist in the model, the model judgment module 1 writes the new parts of the input human-machine hybrid augmented decision model into the corresponding case-based reasoning module library, rule-based reasoning module library, and neural network-based reasoning module library in the cloud decision model library for storage, thereby realizing the updating and optimization of the cloud decision model library. At the same time, the model judgment module 2 of the cloud decision model library management system calls all human-machine hybrid augmented decision models of human-machine co-driving vehicles connected to the cloud and compares the contents of the cloud decision model library with them. When the driving case library, driving rule library, or neural network in the reasoning module of a human-machine hybrid augmented decision model of a human-machine co-driving vehicle is different from that in the cloud decision model library, the reasoning module in the cloud decision model library is downloaded to the human-machine hybrid augmented decision model of that human-machine co-driving vehicle, thereby realizing the updating and online optimization of the human-machine hybrid augmented decision model.

[0069] The specific process of online optimization of the human-machine hybrid enhanced decision prediction model in the third step is as follows:

[0070] The human-machine hybrid augmented decision prediction model adopts a model predictive control-based approach. This method can predict the future dynamics of the system based on the current human-machine hybrid augmented decision information. A kinematic and dynamic model of the human-machine co-driving vehicle is established to reflect the dependence of the system output on current measured variables and current and future inputs. Assuming the system has n state variables x, m input variables u, and p output variables y, the decision prediction model is as follows:

[0071] x k+1 =Ax k +Bu k (13)

[0072] y k =Cx k +Du k (14)

[0073] Where A, B, C, and D are state-space matrices, and x k+2 Use x k+1 and u k+1 This means that we can obtain:

[0074] x k+2 =Ax k+1 +Bu k+1 =Ax k +Bu k +Bu k+1 (15)

[0075] y k+1 =Cx k+1 +Du k+1 =CAx k +CBu k +Du k+1 (16)

[0076] Similarly, the states and outputs of other future time steps are obtained, and the resulting time series is expressed in the form of a block matrix to predict the system from k+1 to k+n. p The time step status is as follows:

[0077]

[0078] The system from k to k+n p The output of time step -1 can be represented as:

[0079]

[0080] The entire prediction process can be completed with just one input time series and one initial state vector.

[0081] Using the time-domain rolling principle, a finite-time-domain optimization decision problem is solved within each sampling period. The solution to this problem is the result of predicting a time-domain length of n. p Given the optimal decision input, the first element of the optimal decision sequence is applied to the system, and the remaining elements are discarded. The system state is then updated, and this process is repeated n times. p The prediction of the time domain length yields the optimal decision sequence. Repeating the above process will give the prediction result of the future dynamics of the system.

[0082] The online optimization of the human-machine hybrid augmented decision prediction model employs a genetic algorithm to optimize the model's parameters. By analyzing the mechanism of the human-machine hybrid augmented decision prediction model, the parameters requiring optimization are identified. Then, an optimization objective function is established, using the standard deviation of the prediction error in the prediction time domain as the objective function. The expression for the objective function is as follows:

[0083]

[0084] Among them, y k+i For the actual output, y' k+i To predict the output, q min q max r min r max These are the upper and lower limits of the weight coefficients q and r, respectively;

[0085] The optimal parameters are obtained by solving the objective function of parameter optimization using a genetic algorithm. Initialization is performed first, setting the number of individuals N in the population and the crossover probability P. c Probability of mutation P m The maximum number of generations (Gen) and the maximum number of generations (Gco) are determined. Then, an initial population is formed by randomly initializing N sets of parameters within the feasible region of the parameters, encoding them using floating-point numbers to obtain a population containing N individuals. The fitness values ​​are then calculated by substituting each of the N individuals into the fitness function, and the individuals are sorted in order of size. The fitness function is shown in the following equation:

[0086]

[0087] Then, a certain number of individuals are selected to form an intermediate population. The probability of an individual being selected into the intermediate population is shown in the following formula:

[0088]

[0089] Then, two individuals are randomly selected from the intermediate population for crossover. The crossover position is randomly chosen with a crossover probability P. c Two new individuals are generated by swapping crossover segments; with mutation probability P m After crossover, mutate each gene locus of the two new individuals in the solution space, and place the mutated individuals into a newly created population; repeat the operations of "selecting individuals," "crossover," and "mutating" until the number of individuals in the new population reaches N, then replace the old population with the new population; repeat the steps after initialization until the stopping condition is met. The stopping condition is:

[0090]

[0091] Where g is the current optimization algebra, X best The best individual generated;

[0092] Finally, the individual with the best fitness in the population is selected for decoding to obtain the optimal parameter combination. The obtained optimal parameters are then put into the human-machine hybrid augmented decision prediction model, thus completing the online optimization of the human-machine hybrid augmented decision prediction model.

[0093] The beneficial effects of this invention are:

[0094] This invention provides an online optimization method for human-machine decision-making logic in highly automated driving, avoiding database limitations, algorithmic limitations, and security risks inherent in human-machine hybrid augmented decision-making processes. It overcomes the technical bottlenecks posed by the uninterpretable and unexplainable characteristics of each level of the human-machine hybrid augmented decision-making model to the safety and reliability of intelligent vehicles, thereby improving the reliability of human-machine hybrid augmented decision-making results and the safety of human-machine co-driving vehicles. Specific beneficial effects are as follows:

[0095] This invention provides an online optimization method for a human-machine hybrid augmented decision database. The single-vehicle self-learning evolution optimization method expands the driving scenarios and situations of human-machine co-driving vehicles, providing a basis for human-machine hybrid augmented decision-making and improving driving safety. The cloud download and update optimization method uploads all new driving scenarios and situations encountered by human-machine co-driving vehicles to the cloud and then updates them synchronously to each vehicle, maximizing the improvement of the human-machine hybrid augmented decision database and giving full play to the advantages of vehicle network information sharing.

[0096] This invention provides an online optimization method for a human-machine hybrid augmented decision-making model. The single-vehicle self-learning evolution optimization method optimizes each level of the decision-making model online by expanding the vehicle's decision database. This includes expanding the driving case library and driving experience library of the decision-making model, improving the neural network neuron structure and internal model parameters, and deriving the model structure to ensure the security of the decision results. The cloud-based download and update optimization method uploads the updated decision models of all human-machine co-driving vehicles to the cloud and then synchronously updates them to each vehicle, maximizing the improvement of the human-machine hybrid augmented decision-making model and fully leveraging the advantages of vehicle network information sharing.

[0097] This invention provides an online optimization method for a human-machine hybrid augmented decision prediction model. By optimizing the parameters of the human-machine hybrid augmented decision prediction model online, the prediction effect on the future state of human-machine co-driving vehicles is greatly improved, ensuring the safety of human-machine co-driving vehicles. Attached Figure Description

[0098] Figure 1 This is a schematic diagram of the overall steps of the online optimization method for human-machine decision-making logic described in this invention.

[0099] Figure 2 This is an exemplary architecture block diagram of step one of the first steps described in this invention.

[0100] Figure 3 This is a structural diagram of the human-machine hybrid enhanced decision database in step one of the first steps of the present invention.

[0101] Figure 4This is an exemplary architecture block diagram of step two of the first step described in this invention.

[0102] Figure 5 This is an exemplary architecture block diagram of step one of the second step described in this invention.

[0103] Figure 6 This is an exemplary architecture block diagram of step one, section two, of the second step described in this invention.

[0104] Figure 7 This is an exemplary architecture block diagram of step two of the second step described in this invention.

[0105] Figure 8 This is an exemplary architecture block diagram of the third step described in this invention. Detailed Implementation

[0106] Please see Figures 1 to 8 As shown:

[0107] The online optimization method for human-machine decision-making logic for highly automated driving provided by this invention is described in detail below:

[0108] Step 1: Online optimization of the human-machine hybrid decision-making database;

[0109] Step 2: Online optimization of the human-machine hybrid augmented decision-making model;

[0110] The third step is to optimize the human-machine hybrid enhanced decision prediction model online.

[0111] The process of online optimization of the human-machine hybrid augmented decision database in the first step is as follows:

[0112] Step 1: Update the human-machine hybrid augmented decision database by sensing new driving situations or scenarios through sensors. The sensor-perceived information is processed and fused, and the knowledge acquisition module converts it into relevant knowledge content. The knowledge judgment module then determines whether the knowledge content belongs to a new knowledge type. If so, the human-machine hybrid augmented decision database is updated for online optimization.

[0113] Step 2: Update the human-machine hybrid augmented decision-making database via cloud download. The updated human-machine hybrid augmented decision-making database for each human-machine co-driving vehicle connected to the cloud is uploaded to the cloud in real time, enriching the cloud database content. Then, the knowledge judgment module updates the content of the cloud database to all human-machine co-driving vehicles.

[0114] In the appendix Figure 2The diagram illustrates an exemplary implementation of step one of the first steps. The human-machine hybrid augmented decision database is part of the human-machine hybrid augmented decision knowledge base management system, and mainly includes a rule base, a meta-knowledge base, a professional vocabulary base, and a multi-level knowledge unit base. The rule base stores various relevant rules and regulations such as traffic rules and driving rules for human-machine co-driving vehicles; the meta-knowledge base stores the priority, category, and macro-level description of knowledge, thereby facilitating knowledge manipulation; the professional vocabulary base stores various professional terms related to human-machine co-driving vehicles, such as traffic-related terms, driving behavior terms, and driver-related terms; the multi-level knowledge unit base stores knowledge information at various levels. The composition of the human-machine hybrid augmented decision database is shown in the attached diagram. Figure 3 As shown, when the knowledge content in the human-machine hybrid augmented decision database is too limited, there may be no solution to the decision or the retrieved decision solutions may be difficult to adapt to new driving situations and scenarios. Therefore, the human-machine hybrid augmented decision database needs to have the function of online updating and optimization to lay the foundation and provide a basis for safe and reliable human-machine hybrid augmented decision-making.

[0115] Updating the database by sensing new driving situations or scenarios through sensors is a type of database optimization method based on single-vehicle self-learning evolution. When new driving situations or scenarios arise in the driving environment of a human-machine co-driving vehicle, the decision-making capabilities of the human-machine hybrid augmented decision-making system often exceed its limits. Therefore, to ensure the safety of human-machine hybrid augmented decision-making, it is necessary to record the new driving situations and scenarios, convert them into new knowledge content, and store them in the human-machine hybrid augmented decision-making database to update and enrich the database content and provide a basis for human-machine hybrid augmented decision-making.

[0116] Human-machine co-driving vehicles perceive their own position, status, and information about the surrounding environment through various types of sensors. These sensors can be categorized into three types: positioning sensors, vision sensors, and radar sensors. Among positioning sensors, the Global Positioning System (GPS) can provide high-precision position information for human-machine co-driving vehicles. However, its positioning effectiveness depends on satellite signals and is susceptible to environmental interference. In urban environments with numerous buildings and dense vegetation, GPS alone is insufficient for navigation. Therefore, a combined navigation system using GPS and an inertial measurement unit (IMU) has become a trend. This system fully utilizes the information collected by both GPS and IMU sensors. High-precision GPS positioning results are used to control the system drift of the IMU, reducing the accumulation of measurement errors. Simultaneously, an encoder measures the travel distance of the human-machine co-driving vehicle, enabling position and absolute pose estimation, and outputting the vehicle's position and pose information. Vision sensors acquire image information of the surrounding environment from vision cameras. By analyzing the pixel attributes of the images, the system infers the attributes and status of objects in the environment. For example, cameras can identify traffic participants in the vehicle's surrounding environment, outputting their type and status information; they can identify traffic signs, outputting their type and content information; and they can compare and judge street scenes with navigation maps to output the vehicle's precise location information. One drawback of visual sensors is their susceptibility to lighting conditions, leading to significant errors in areas with indistinct features, such as open roads. Radar sensors, such as lidar and millimeter-wave radar, can measure the status of dynamic and static traffic participants around the human-machine co-driving vehicle, outputting information such as the participants' speed, distance, and orientation.

[0117] After acquiring environmental information about the surroundings of the human-machine co-driving vehicle, various types of sensors need to process and combine the information collected by multiple sensors. This enables the vehicle's perception layer to sense the surrounding environment with low false alarm rate and high output efficiency, thereby reducing blind spots, improving detection accuracy, simplifying data output dimensions, and providing a concise and reliable data foundation for further functional decisions, thus ensuring the driving safety of the human-machine co-driving vehicle. Sensor fusion consists of three parts: time synchronization, spatial synchronization, and data processing. Time synchronization filters information collected by different sensors within a specified time period and sorts it according to accuracy using methods such as interpolation and extrapolation, thereby achieving uniformity among different sensors at the same time. Spatial synchronization solves the coordinate transformation relationship between different sensors through the calibration of individual sensors and joint calibration of multiple sensors. After spatial synchronization, it maps the information of the same object under different sensors, thus solving the information collected by different sensors into the same coordinate system. Data processing addresses the parts with large differences, data redundancy, and missing data in the synchronized sensor measurements to achieve data compatibility among sensors. Then, it processes the data of the same object measured by different sensors, selecting the most accurate measurement value from multiple measurements as the sensor output, simplifying the data output dimension.

[0118] The information fused from the sensors is input into the knowledge acquisition module of the human-machine hybrid augmented decision-making knowledge base management system. This module converts the sensor-fused information about the human-machine co-driving vehicle and its surrounding environment into driving knowledge terminology, driving knowledge experience rules, driving knowledge rule knowledge units, and behavioral decision meta-knowledge units, and then inputs this knowledge into the knowledge judgment module. The knowledge judgment module can call upon knowledge from the rule base, meta-knowledge base, terminology base, and multi-level knowledge unit base in the current human-machine hybrid augmented decision-making database and compare it with the knowledge input from the knowledge acquisition module. When new knowledge types exist (such as knowledge about the vehicle itself, surrounding vehicles, road environment, and traffic rules), the knowledge judgment module writes the new knowledge content into the corresponding terminology base, experience rule base, rule knowledge unit base, and meta-knowledge unit base in the human-machine hybrid augmented decision-making database according to its category, thus achieving the updating and optimization of the human-machine hybrid augmented decision-making database.

[0119] In the appendix Figure 4The diagram illustrates an exemplary implementation of step two of the first step. Unlike the single-vehicle self-learning evolutionary optimization database, updating the database via the cloud allows human-machine co-driving vehicles to optimize their own human-machine hybrid decision-making database by learning the driving situations and scenarios of other vehicles. Since the driving situations and scenarios a vehicle encounters in a real road environment are limited, and the greatest advantage of intelligent connected vehicles lies in vehicle-to-vehicle connectivity and information sharing, updating the database via the cloud can greatly expand the content of the human-machine hybrid augmented decision-making database.

[0120] Human-machine co-driving vehicles are intelligent connected vehicles that connect to the cloud via a network. They can upload their own human-machine hybrid decision-making knowledge base to the cloud database in real time, and also download and update their own human-machine hybrid augmented knowledge base from the cloud database in real time. When encountering a new driving situation or scenario, each human-machine co-driving vehicle first updates its own human-machine hybrid augmented decision-making database through sensor-based database updates. After the update is complete, the vehicle promptly uploads the updated database content to the cloud via the network. Upon receiving the new knowledge content, the cloud inputs it into the knowledge judgment module 1 of the cloud knowledge base management system. The knowledge judgment module 1 can call upon knowledge from the rule base, meta-knowledge base, professional vocabulary base, and multi-level knowledge unit base in the current cloud database and compare it with the newly input knowledge content. When a new knowledge type exists (such as knowledge about the vehicle itself, surrounding environment vehicles, road environment, and traffic rules), the knowledge judgment module 1 writes the new knowledge content into the corresponding professional vocabulary base, experience rule base, rule knowledge unit base, and meta-knowledge unit base in the cloud database according to its category, thus achieving the updating and optimization of the cloud database. At the same time, the knowledge judgment module 2 of the cloud database knowledge base management system will call the human-machine hybrid augmented decision database of all human-machine co-driving vehicles connected to the cloud and compare the contents of the cloud database with it. When the human-machine hybrid augmented decision database of a certain human-machine co-driving vehicle is missing a certain knowledge in the cloud database, the knowledge will be downloaded to the human-machine hybrid augmented decision database of that human-machine co-driving vehicle to realize the updating and optimization of the human-machine hybrid augmented decision database.

[0121] The second step involves online optimization of the human-machine hybrid augmented decision-making model as follows:

[0122] Step 1: Optimize the human-machine hybrid augmented decision-making model online by consulting with drivers. This involves constructing new knowledge inferences between the updated knowledge in the human-machine hybrid augmented decision-making database and the decision results. These new knowledge inferences are then used to optimize the human-machine hybrid augmented decision-making model online, achieving incremental learning of the database. Step 1 consists of two parts.

[0123] Step 1: Evaluating the Model's Decision-Making Capability. After receiving sensory input, the human-machine hybrid decision-making model progressively enters different reasoning modules within the model. Each module then determines whether it has the capability to output a decision based on the sensory input. If all reasoning modules in the model are unable to provide knowledge-based reasoning for behavioral decisions based on the sensory input, it indicates that the model's decision-making capability is poor and it cannot handle the current sensory input situation. In this case, the model should consult a human driver for appropriate action.

[0124] Step Two: Online Model Optimization. After the human driver gives a driving command, the human-machine hybrid augmented decision model records the human driver's driving operation. By learning the perceptual input and the corresponding human driver's driving operation, it establishes decision knowledge reasoning between perceptual input and decision output, thereby optimizing the case-based reasoning module, rule-based reasoning module, and neural network-based reasoning module in the decision model online.

[0125] Step 2: Optimize the human-machine hybrid augmented decision-making model online via cloud download. By uploading the updated human-machine hybrid augmented decision-making model of each human-machine co-driving vehicle connected to the cloud in real time, the content of the cloud decision model library is enriched. Then, the model judgment module updates the content of the cloud decision model library to all human-machine co-driving vehicles. This cloud-based decision model update method allows human-machine co-driving vehicles to optimize their own human-machine hybrid augmented decision-making model by learning the knowledge reasoning between the perception inputs and decision outputs of other vehicles.

[0126] In the appendix Figure 5 The diagram illustrates an exemplary implementation of step one of the first steps in the second step. Knowledge reasoning is a crucial step in transforming the perceptual input of a human-machine co-driving vehicle into decision output; the efficiency and reliability of knowledge reasoning significantly impact the effectiveness of the decision. The human-machine hybrid augmented decision-making online optimization model achieves incremental learning of the human-machine hybrid augmented decision-making knowledge base through online optimization using knowledge reasoning.

[0127] The human-machine hybrid augmented decision-making model includes a case-based reasoning module, a rule-based reasoning module, and a neural network-based reasoning module. The sensor-fused input includes information such as the vehicle's driving state, the motion state of surrounding traffic participants, traffic signs, road conditions, weather conditions, and lighting conditions. This information, after being input into the human-machine hybrid augmented decision-making model, first enters the case-based reasoning module. Upon receiving the input, the case-based reasoning module searches the stored driving case library and calculates the maximum similarity between the perceived input and the stored cases in the library. The calculation formula is as follows:

[0128]

[0129] Among them, f ij ω represents the similarity between the j-th attribute of the i-th case in the case library and the j-th attribute of the input case. j F represents the weight coefficient of the j-th attribute. i This represents the similarity between the input case and the i-th case.

[0130] When the maximum similarity is below a threshold, it indicates that the case-based reasoning module cannot generate an effective decision output for the current perceived input. The perceived input is then passed to the next-level rule-based reasoning module. Upon receiving the input, the rule-based reasoning module compares the perceived input with knowledge in the driving rule base. If an unknown condition exists, it indicates that the rule-based reasoning module cannot generate an effective decision output for the current perceived input. The perceived input is then passed to the next-level neural network-based reasoning module. Upon receiving the input, the neural network-based reasoning module determines whether the perceived input is within the range of values ​​for the neural network's input variables. If the perceived input exceeds the range of values ​​for the neural network's input variables, it indicates that the neural network-based reasoning module cannot generate an effective decision output for the current perceived input. The module then consults the human driver for appropriate action.

[0131] In the appendix Figure 6 The diagram illustrates an exemplary implementation of step one, section two of the second step. The case-based reasoning module performs online optimization by storing the perceptual input and the corresponding human driver's driving operations as new driving cases in a driving case library.

[0132] The rule-based reasoning module performs online optimization by storing perceptual inputs and corresponding human driver actions as new driving rules and driving experiences in the driving rule base and driving experience base.

[0133] The neural network-based reasoning module uses a deep learning-based neural network for knowledge reasoning. It inputs fused information from the sensors of the human-machine co-driving vehicle into the neural network model, and through the model's computation, provides accurate decision-making information. After obtaining new knowledge reasoning between the perceptual input and the corresponding driving operations of the human driver, the neural network-based reasoning module uses an evolutionary algorithm to adjust the brain-like neuron structure, internal model parameter dimensions, and organizational form within the model, deriving the model structure and completing online optimization.

[0134] The basic computational unit of a neural network is the neuron, and the formula for calculating the output is as follows:

[0135] C=f(XW+b) (2)

[0136] Where X is the input signal of each layer, W is the weight corresponding to the input signal, b is the bias, which can enhance the expressive ability of the neuron, f is the activation function, and C is the output signal of each layer.

[0137] A fully connected neural network is trained by continuously adjusting its weights and biases in response to perceived input stimuli, so that the network's output consistently approximates the desired decision outcome. During neural network training, a non-linear, thresholded activation function, the Sigmoid function, is chosen, and its calculation formula is as follows:

[0138]

[0139] The formula for calculating the output of the nth hidden layer in a neural network is:

[0140]

[0141] Where x m Let ω be the input vector. nm Let b be the weights corresponding to nodes from the m-th layer to the n-th layer, where N is the number of nodes in the n-th layer. n For the bias of the nth layer, Hide n It is the output of the nth layer.

[0142] The formula for calculating the output layer is:

[0143]

[0144] Where C is the output signal of the output layer, ω cn Here, b represents the weights of the nth hidden layer, P is the total number of hidden layers, and b is the weights of the nth hidden layer. c For the bias of the output layer, f c This is the activation function for the output layer.

[0145] The residual E is obtained by calculating the mean square error between the output C and the actual label. The calculation formula is as follows:

[0146]

[0147] The weights from the output layer to the hidden layer are updated using gradient descent. The impulse of the weights and biases is calculated using the chain rule, and then both are updated separately. The formulas for calculating the impulse of the weights and biases are as follows:

[0148]

[0149]

[0150] in and Here, η represents the impulse of the weights and biases, respectively; η is the learning rate; and O is the intermediate function.

[0151] Finally, update all weights and biases between the hidden layer and the input layer layer by layer. The formula for calculating the impulse of the weights and biases is as follows:

[0152]

[0153]

[0154] in and These are the impulses for the weights and biases, respectively.

[0155] Evolutionary algorithms encode various parameters of a neural network into chromosomes, and then use these chromosomes to build the network and optimize the parameters. This type of evolutionary neural network employs a weighted topological evolution method, which can automatically evolve a suitable neural network structure based on perceived input and the human driver's decision output. This structure includes network weights, network structure, and activation functions.

[0156] The weighted topology evolution method employs an evolutionary neural network based on enhanced topology. First, a complete neural network model, including the network topology and connection weights, is constructed and encoded as node genes and weight genes.

[0157] Then, mutation and crossover operations are performed, and an innovation ID is introduced to record the history of gene crossover and mutation. The mutation operation includes four parts: adding a node, adding connections between nodes, adjusting weights, and changing the activation function response. The crossover operation follows the principle of "matching gene points are used directly in the child gene, disconnected gene points or excess genes are obtained from the parent gene with high fitness, and two parent genes with the same fitness are randomly inherited."

[0158] Population dissimilarity distance is defined using disjointed gene points, matched gene points, and excess gene points between two genes. The more mismatches between two genes, the greater the structural difference between them. The calculation formula is:

[0159]

[0160] Where d represents the difference distance, N represents the number of excess gene points in the two genes, M represents the number of disjoint gene points in the two genes, a1, a2, and a3 are coefficients, P is the average of the weight differences of the matching gene points, and L is the length of the longer gene of the two genes. Based on this formula, networks with similar structures and weights can be assigned to the same group, allowing for intra-group competition during evolution. This method prevents populations with new structures from immediately disappearing, giving them sufficient time to adapt to the environment.

[0161] To prevent a few genes in a group from exhibiting excessively high fitness to the environment, thus masking the performance of other genes with lower fitness and reducing overall species diversity, explicit fitness is defined. This means that genes belonging to the same group share the same level of fitness to the environment. The formula for calculating explicit fitness is:

[0162]

[0163] Where f i denoted by δ(i,j), represents the genetic fitness, where δ(i,j) is the difference between two individuals in the group. When δ(i,j) is less than the difference value, sh(δ(i,j)) is 1, and when δ(i,j) is greater than the difference value, sh(δ(i,j)) is 0. n is the total number of individuals in the group.

[0164] By sharing fitness, each gene is constrained. When there are too many members in a group, its fitness rating is lowered; and when a new gene is generated, it is protected. In this way, if an older gene does not improve its environmental fitness after multiple generations of evolution, it will gradually be eliminated.

[0165] The evolutionary algorithm trains the deep learning-based neural network in the neural network-based reasoning module based on the new knowledge reasoning formed between the perceptual input and the human driver's decision output. It adjusts the internal brain-like neuron structure, internal model parameter dimensions and organization form to derive the model structure, realize incremental learning of the human-machine hybrid augmented decision database, and complete the online optimization of the neural network-based reasoning module.

[0166] In the appendix Figure 7 The diagram illustrates an exemplary implementation of step two of the second step. Unlike the single-vehicle self-learning evolutionary optimization of the human-machine hybrid augmented decision-making model, updating the decision-making model via the cloud allows the human-machine co-driving vehicle to optimize its own human-machine hybrid augmented decision-making model by learning the knowledge reasoning between the perception inputs and decision outputs of other vehicles. Since the driving situations and scenarios a vehicle encounters in real-world road environments are limited, and the greatest advantage of intelligent connected vehicles lies in vehicle-to-vehicle connectivity and information sharing, updating the human-machine hybrid augmented decision-making model via the cloud can greatly improve the vehicle's human-machine hybrid augmented decision-making model, enhancing the safety and reliability of decision-making.

[0167] Human-machine co-driving vehicles are intelligent connected vehicles. Connected to the cloud via a network, they can upload their own human-machine hybrid augmented decision-making model to the cloud decision model library in real time, and also download and update their own human-machine hybrid augmented decision-making model from the cloud decision model library in real time. Each human-machine co-driving vehicle, after generating new perceptual inputs and decision outputs, and completing online optimization of its case-based reasoning module, rule-based reasoning module, and neural network-based reasoning module, will promptly upload its newly updated human-machine hybrid augmented decision-making model to the cloud via the network. Upon receiving the input human-machine hybrid augmented decision-making model, the cloud will input it into the model judgment module 1 of the cloud decision model library management system. Model Judgment Module 1 can compare the case-based, rule-based, and neural network-based reasoning module libraries in the current cloud-based decision model library with the corresponding modules of the input decision model. When a new driving case library, driving rule library, or neural network exists in the input decision model, Model Judgment Module 1 writes the new part of the input human-machine hybrid augmented decision model into the corresponding case-based, rule-based, and neural network-based reasoning module libraries in the cloud-based decision model library for storage, thereby updating and optimizing the cloud-based decision model library. Simultaneously, Model Judgment Module 2 of the cloud-based decision model library management system calls all human-machine hybrid augmented decision models of human-machine co-driving vehicles connected to the cloud and compares them with the contents of the cloud-based decision model library. When the driving case library, driving rule library, or neural network in the reasoning module of a human-machine hybrid augmented decision model of a human-machine co-driving vehicle differs from that in the cloud-based decision model library, the corresponding reasoning module in the cloud-based decision model library is downloaded to the human-machine hybrid augmented decision model of that human-machine co-driving vehicle, thereby updating and optimizing the human-machine hybrid augmented decision model online.

[0168] The third step is to perform online optimization of the human-machine hybrid augmented decision prediction model. The parameters of the human-machine hybrid augmented decision prediction model are optimized using a genetic algorithm. First, the model parameters to be optimized are determined, and an objective function for parameter optimization is established. Then, the optimal solution is searched in the solution space using a genetic algorithm until the population converges. Finally, the optimal individual is decoded to obtain the parameter combination, which is then used to replace the parameters in the model, completing the online optimization of the human-machine hybrid augmented decision prediction model.

[0169] In the appendix Figure 8An exemplary implementation of the third step is shown. Based on a rolling time-domain online human-machine hybrid augmented decision prediction model, it can predict the impact of human-machine hybrid augmented decision results on the future state of the human-machine co-driving vehicle. As the human-machine hybrid augmented decision database and the human-machine hybrid augmented decision model are optimized online, the human-machine hybrid augmented decision prediction model should also be updated accordingly. Therefore, establishing an online evaluation model for the human-machine hybrid augmented decision prediction model and optimizing its parameters online can greatly improve the prediction effect on the future state of the human-machine co-driving vehicle, which is beneficial to the safety and reliability of human-machine hybrid augmented decision-making.

[0170] The human-machine hybrid augmented decision prediction model employs a model predictive control (MMCC) approach, which can predict the future dynamics of the system based on current human-machine hybrid augmented decision information. A kinematic and dynamic model of the human-machine co-driving vehicle is established to reflect the dependence of the system output on current measured variables and current and future inputs. Assuming the system has n state variables x, m input variables u, and p output variables y, the decision prediction model can be expressed as:

[0171] x k+1 =Ax k +Bu k (13)

[0172] y k =Cx k +Du k (14)

[0173] Where A, B, C, and D are state-space matrices. Let x... k+2 Use x k+1 and u k+1 This means that we can obtain:

[0174] x k+2 =Ax k+1 +Bu k+1 =Ax k +Bu k +Bu k+1 (15)

[0175] y k+1 =Cx k+1 +Du k+1 =CAx k +CBu k +Du k+1 (16)

[0176] Similarly, the states and outputs of other future time steps can be obtained. Therefore, the obtained time series can be expressed in the form of a block matrix to predict the system from k+1 to k+n. p The state of a time step can be represented as:

[0177]

[0178] The system from k to k+n p The output of time step -1 can be represented as:

[0179]

[0180] By expressing it this way, predicting future system changes becomes very intuitive, requiring only an input time series and an initial state vector to complete the entire prediction process.

[0181] Using the time-domain rolling principle, a finite-time-domain optimization decision problem is solved within each sampling period. The solution to this problem is the result of predicting a time-domain length of n. p The optimal decision input is then determined. Next, the first element of the optimal decision sequence is applied to the system, and the remaining elements are discarded. The system state is updated, and the process is repeated n times. p The optimal decision sequence is obtained by predicting the time domain length. Repeating the above process yields the prediction result of the system's future dynamics.

[0182] The online optimization of the human-machine hybrid augmented decision prediction model employs a genetic algorithm to optimize the model's parameters. By analyzing the mechanism of the human-machine hybrid augmented decision prediction model, the parameters requiring optimization are identified, and then an optimization objective function is established. The standard deviation of the prediction error within the prediction time domain can be used as the objective function, and its expression is as follows:

[0183]

[0184] Among them, y k+i For the actual output, y' k+i To predict the output, q min q max r min r max These are the upper and lower limits of the weight coefficients q and r, respectively.

[0185] The optimal parameters are obtained by solving the objective function of parameter optimization using a genetic algorithm. Initialization is performed first, setting the number of individuals N in the population and the crossover probability P. c Probability of mutation P m The maximum number of generations (Gen) and the maximum number of generations (Gco) are determined. Then, an initial population is formed by randomly initializing N sets of parameters within the feasible region of the parameters, encoding them using floating-point numbers to obtain a population containing N individuals. The fitness values ​​are then calculated by substituting each of the N individuals into the fitness function, and the individuals are sorted in order of size. The fitness function is shown in the following equation:

[0186]

[0187] Then, a certain number of individuals are selected to form an intermediate population. The probability of an individual being selected into the intermediate population is shown in the following formula:

[0188]

[0189] Then, two individuals are randomly selected from the intermediate population for crossover. The crossover position is randomly chosen with a crossover probability P. c Two new individuals are generated by swapping crossover segments; with mutation probability P m After crossover, mutate each gene locus of the two new individuals in the solution space, and place the mutated individuals into a newly created population; repeat the operations of "selecting individuals," "crossover," and "mutating" until the number of individuals in the new population reaches N, then replace the old population with the new population; repeat the steps after initialization until the stopping condition is met. The stopping condition is:

[0190]

[0191] Where g is the current optimization algebra, X best The best individual generated.

[0192] Finally, the individual with the best fitness in the population is selected for decoding to obtain the optimal parameter combination. The obtained optimal parameters are then fed into the human-machine hybrid augmented decision prediction model, thus completing the online optimization of the model.

Claims

1. An online optimization method for human-machine decision-making logic in highly automated driving, characterized in that: The method includes the following steps: The first step is to optimize the online decision-making database using a human-machine hybrid approach. The specific steps are as follows: Step 1: Update the human-machine hybrid augmented decision database by sensing new driving situations or scenarios through sensors. Process and fuse the sensor information, convert it into relevant knowledge content by the knowledge acquisition module, and then determine whether the knowledge content belongs to a new knowledge type by the knowledge judgment module. If so, update the human-machine hybrid augmented decision database for online optimization. Step 2: Update the human-machine hybrid augmented decision database by downloading from the cloud. The updated human-machine hybrid augmented decision database of each human-machine co-driving vehicle connected to the cloud is uploaded to the cloud in real time to enrich the content of the cloud database. Then, the knowledge judgment module updates the content of the cloud database to all human-machine co-driving vehicles. Step 2: Online optimization of the human-machine hybrid augmented decision-making model; Step 1: Optimize the human-machine hybrid augmented decision-making model online by consulting with drivers. Construct new knowledge reasoning between the updated knowledge in the human-machine hybrid augmented decision-making database and the decision results. Then, use this new knowledge reasoning to optimize the human-machine hybrid augmented decision-making model online, achieving incremental learning of the human-machine hybrid augmented decision-making database. Step 1 is completed in two stages, as follows: Step 1: Evaluating the Model's Decision-Making Capability: After receiving sensory input, the human-machine hybrid decision-making model will pass through different reasoning modules in the human-machine hybrid augmented decision-making model. Each reasoning module will determine in turn whether it has the ability to make a decision output for the sensory input. When all reasoning modules in the human-machine hybrid augmented decision-making model are unable to provide knowledge reasoning for making a behavioral decision based on the sensory input, it indicates that the current human-machine hybrid augmented decision-making model has poor decision-making capability and cannot cope with the current sensory input situation. In this case, the model will turn to the human driver for consultation, and the human driver will provide specific instructions. Step 2: Online Model Optimization: After the human driver gives the driving operation, the human-machine hybrid augmented decision model records the human driver's driving operation and establishes decision knowledge reasoning between the perceptual input and the corresponding human driver's driving operation by learning the perceptual input and the decision output. This allows for online optimization of the case-based reasoning module, rule-based reasoning module, and neural network-based reasoning module in the decision model. Step 2: Optimize the human-machine hybrid augmented decision-making model online via cloud download. This involves uploading the updated human-machine hybrid augmented decision-making model of each human-machine co-driving vehicle connected to the cloud in real time, enriching the cloud decision model library. Then, the model judgment module updates the content of the cloud decision model library to all human-machine co-driving vehicles. By updating the decision model through the cloud, the human-machine co-driving vehicles can optimize their own human-machine hybrid augmented decision-making model by learning the knowledge reasoning between the perception input and decision output of other vehicles. Step 3: Online optimization of the human-machine hybrid augmented decision prediction model: The parameters of the human-machine hybrid augmented decision prediction model are optimized by means of a genetic algorithm. First, the model parameters to be optimized are determined and the objective function for parameter optimization is established. Then, the optimal solution is searched in the solution space by the genetic algorithm until the population converges. Finally, the optimal individual is decoded to obtain the parameter combination, which is used to replace the parameters in the model to complete the online optimization of the human-machine hybrid augmented decision prediction model. The specific process of online optimization of the human-machine hybrid augmented decision prediction model is as follows: The human-machine hybrid augmented decision prediction model adopts a model predictive control-based approach. This method can predict the future dynamics of the system based on the current human-machine hybrid augmented decision information. A kinematic and dynamic model of the human-machine co-driving vehicle is established to reflect the dependence of the system output on current measured variables and current and future inputs. The system has n state variables x, m input variables u, and p output variables y. The decision prediction model is as follows: （13） （14） Where A, B, C, and D are state-space matrices, and x k+2 Use x k+1 and u k+1 This means that we can obtain: （15） （16） Similarly, the states and outputs of other future time steps are obtained, and the resulting time series is expressed in the form of a block matrix to predict the system from k+1 to k+n. p The time step status is as follows: （17） The system from k to k+n p The output of time step -1 is represented as: （18） The entire prediction process can be completed with just one input time series and one initial state vector. Using the time-domain rolling principle, a finite-time-domain optimization decision problem is solved within each sampling period. The solution to this problem is the result of predicting a time-domain length of n. p Given the optimal decision input, the first element of the optimal decision sequence is applied to the system, and the remaining elements are discarded. The system state is then updated, and this process is repeated n times. p The prediction of the time domain length yields the optimal decision sequence. Repeating the above process will give the prediction result of the future dynamics of the system. The online optimization of the human-machine hybrid augmented decision prediction model employs a genetic algorithm to optimize the model's parameters. By analyzing the mechanism of the human-machine hybrid augmented decision prediction model, the parameters requiring optimization are identified. Then, an optimization objective function is established, using the standard deviation of the prediction error in the prediction time domain as the objective function. The expression for the objective function is as follows: （19） Among them, y k+i For the actual output, To predict the output, q min q max r min r max These are the upper and lower limits of the weight coefficients q and r, respectively; The optimal parameters are obtained by solving the objective function of parameter optimization using a genetic algorithm. Initialization is performed first, setting the number of individuals N in the population and the crossover probability P. c Probability of mutation P m The maximum number of generations (Gen) and the maximum number of generations (Gco) are determined. Then, an initial population is formed by randomly initializing N sets of parameters within the feasible region of the parameters, encoding them using floating-point numbers to obtain a population containing N individuals. The fitness values ​​are then calculated by substituting each of the N individuals into the fitness function, and the individuals are sorted in order of size. The fitness function is shown in the following equation: （20） Then, a certain number of individuals are selected to form an intermediate population. The probability of an individual being selected into the intermediate population is shown in the following formula: （21） Then, two individuals are randomly selected from the intermediate population for crossover. The crossover position is randomly chosen with a crossover probability P. c Two new individuals are generated by swapping crossover segments; with mutation probability P m After crossover, mutate each gene locus of the two new individuals in the solution space, and place the mutated individuals into a newly created population. Repeat the operations of "selecting individuals," "crossover," and "mutating" until the number of individuals in the new population reaches N, then replace the old population with the new population. Repeat the steps after initialization until the stopping condition is met. The stopping condition is: （22） Where g is the current optimization algebra, X best The best individual generated; Finally, the individual with the best fitness in the population is selected for decoding to obtain the optimal parameter combination. The obtained optimal parameters are then put into the human-machine hybrid augmented decision prediction model, thus completing the online optimization of the human-machine hybrid augmented decision prediction model. The specific process of step one in the first step is as follows: 1) The Human-Machine Hybrid Augmented Decision Database is part of the Human-Machine Hybrid Augmented Decision Knowledge Base Management System. It includes a rule base, a meta-knowledge base, a professional terminology base, and a multi-level knowledge unit base. The rule base stores various relevant rules and regulations for traffic and driving rules of human-machine co-driving vehicles; the meta-knowledge base stores the priority, category, and macro-level description of knowledge to facilitate knowledge manipulation; the professional terminology base stores various professional terms related to human-machine co-driving vehicles; and the multi-level knowledge unit base stores knowledge information at various levels. If the knowledge content in the Human-Machine Hybrid Augmented Decision Database is too small, there may be no solution to the decision or the retrieved decision solution may be difficult to adapt to new driving situations and scenarios. Therefore, the Human-Machine Hybrid Augmented Decision Database needs to have online update and optimization capabilities to lay the foundation and provide a basis for safe and reliable human-machine hybrid augmented decision-making. 2) Updating the database by sensing new driving situations or scenarios through sensors is a kind of database optimization method of single-vehicle self-learning evolution. When new driving situations or scenarios exist in the driving environment of human-machine co-driving vehicles, the decision-making ability of the human-machine hybrid augmented decision system often exceeds its boundaries. Therefore, in order to ensure the safety of human-machine hybrid augmented decision-making, it is necessary to record the new driving situations and scenarios, convert them into new knowledge content and store them in the human-machine hybrid augmented decision database, update and enrich the database content, and provide a basis for human-machine hybrid augmented decision-making. 3) Human-machine co-driving vehicles perceive their own position, status, and information about the surrounding environment through various types of sensors. These sensors are categorized into three types: positioning sensors, vision sensors, and radar sensors. The positioning sensors fully utilize information collected by GPS and inertial measurement unit (IMU) sensors. High-precision GPS positioning results are used to control the system drift of the IMU, reducing the accumulation of measurement errors. Simultaneously, an encoder measures the vehicle's travel distance, enabling position and absolute pose estimation and outputting the vehicle's position and pose information. Vision sensors acquire image information of the surrounding environment through vision cameras. By analyzing the pixel attributes of the images, they infer the attributes and states of objects in the environment. For example, cameras can identify traffic participants in the surrounding environment and output their type and state information; they can identify traffic signs and output their type and content information; and they can capture street scenes and compare them with navigation maps to output the vehicle's precise location information. Radar sensors, such as lidar and millimeter-wave radar, can measure the state of dynamic and static traffic participants around the human-machine co-driving vehicle and output information about their speed, distance, and orientation. 4) After acquiring environmental information around the human-machine co-driving vehicle, various types of sensors process and combine the information collected by multiple sensors. This enables the vehicle's perception layer to sense the surrounding environment with low false alarm rate and high output efficiency, thereby reducing blind spots, improving detection accuracy, and simplifying data output dimensions. This provides a concise and reliable data foundation for further functional decisions, ensuring the driving safety of the human-machine co-driving vehicle. Sensor fusion consists of three parts: time synchronization, spatial synchronization, and data processing. Time synchronization filters out the information collected by different sensors within a specified time period and sorts them according to accuracy using interpolation and extrapolation methods. This achieves unification of different sensors at the same time; spatial synchronization solves the coordinate transformation relationship between different sensors through the calibration of a single sensor and the joint calibration of several sensors. After spatial synchronization is completed, the information of the same object under different sensors is mapped to the same coordinate system, thereby solving the information collected by different sensors into the same coordinate system; data processing handles the parts with large differences, data redundancy and data missing in the synchronized sensor measurements to achieve data compatibility of each sensor. Then, the data of the same object measured by different sensors are processed, and the measurement value with the highest accuracy is selected as the sensor output from several measurements, simplifying the data output dimension; 5) The information fused from the sensors is input into the knowledge acquisition module of the human-machine hybrid augmented decision-making knowledge base management system. The knowledge acquisition module can convert the information from the sensor-fused human-machine co-driving vehicle and its surrounding environment into knowledge of driving knowledge professional terms, driving knowledge experience rules, driving knowledge rule knowledge units, and behavioral decision meta-knowledge units, and input it into the knowledge judgment module. The knowledge judgment module can call the knowledge in the rule base, meta-knowledge base, professional terminology base, and multi-level knowledge unit base of the current human-machine hybrid augmented decision database and compare it with the knowledge input in the knowledge acquisition module. When there is a new type of knowledge, the knowledge judgment module writes the new knowledge content into the corresponding professional terminology base, experience rule base, rule knowledge unit base, and meta-knowledge unit base of the human-machine hybrid augmented decision database according to its category for storage, so as to realize the updating and optimization of the human-machine hybrid augmented decision database. The process of step two in the first step is as follows: Human-machine co-driving vehicles are intelligent connected vehicles that connect to the cloud via a network. They upload their human-machine hybrid decision-making knowledge base to the cloud database in real time and download and update their human-machine hybrid augmented knowledge base from the cloud database in real time. When encountering a new driving situation or scenario, each human-machine co-driving vehicle first updates its human-machine hybrid augmented decision-making database through sensor-based database updates. After the update is complete, the vehicle promptly uploads the updated database content to the cloud via the network. Upon receiving the new knowledge content, the cloud inputs it into the knowledge judgment module 1 of the cloud knowledge base management system. The knowledge judgment module 1 can access knowledge from the rule base, meta-knowledge base, professional vocabulary base, and multi-level knowledge unit base in the current cloud database. The knowledge judgment module 1 compares the new knowledge content with the newly input knowledge content. When a new knowledge type exists, the knowledge judgment module 1 writes the new knowledge content into the corresponding professional vocabulary library, experience rule library, rule knowledge unit library and meta knowledge unit library in the cloud database according to its category for storage, so as to realize the update and optimization of the cloud database. At the same time, the knowledge judgment module 2 of the cloud database knowledge base management system will call the human-machine hybrid augmented decision database of all human-machine co-driving vehicles connected to the cloud and compare the content of the cloud database with it. When the human-machine hybrid augmented decision database of a certain human-machine co-driving vehicle is missing a certain knowledge in the cloud database, the knowledge is downloaded to the human-machine hybrid augmented decision database of that human-machine co-driving vehicle, so as to realize the update and optimization of the human-machine hybrid augmented decision database. The specific process of step one in step two is as follows: The human-machine hybrid augmented decision-making model includes a case-based reasoning module, a rule-based reasoning module, and a neural network-based reasoning module. The sensor fusion-based perceptual input includes information on the vehicle's driving state, the motion state of surrounding traffic participants, traffic signs, road conditions, weather conditions, and lighting conditions. This information is input into the human-machine hybrid augmented decision-making model and first enters the case-based reasoning module. Upon receiving the input, the case-based reasoning module searches the stored driving case library and calculates the maximum similarity between the perceptual input and the stored cases in the library. The calculation formula is as follows: （1） Among them, f ij ω represents the similarity between the j-th attribute of the i-th case in the case library and the j-th attribute of the input case. j F represents the weight coefficient of the j-th attribute. i This represents the similarity between the input case and the i-th case; When the maximum similarity is below a threshold, it indicates that the case-based reasoning module cannot generate an effective decision output for the current perceived input. The perceived input is then passed to the next-level rule-based reasoning module. Upon receiving the input, the rule-based reasoning module compares it with knowledge in the driving rule base. If an unknown condition exists, it indicates that the rule-based reasoning module cannot generate an effective decision output for the current perceived input. The perceived input is then passed to the next-level neural network-based reasoning module. Upon receiving the input, the neural network-based reasoning module determines whether the perceived input is within the range of values ​​for the neural network's input variables. If the perceived input exceeds the range of values ​​for the neural network's input variables, it indicates that the neural network-based reasoning module cannot generate an effective decision output for the current perceived input. Finally, it consults the human driver for specific instructions.

2. The online optimization method for human-machine decision-making logic for highly automated driving according to claim 1, characterized in that: The specific process of step two in step one of the second step is as follows: 1) The case-based reasoning module performs online optimization by storing perceptual inputs and corresponding human driver operations as new driving cases in the driving case library; 2) The rule-based reasoning module performs online optimization by storing the perceptual input and the corresponding human driver's driving operations as new driving rules and driving experience into the driving rule base and driving experience base. 3) The neural network-based reasoning module performs knowledge reasoning using a deep learning-based neural network. It inputs the information fused from the sensors of the human-machine co-driving vehicle into the neural network model. Through the model's calculations, it provides accurate decision information. After obtaining new knowledge reasoning between the perceptual input and the corresponding driving operations of the human driver, the neural network-based reasoning module adjusts the brain-like neuron structure, internal model parameter dimensions, and organizational form within the model through evolutionary algorithms, deriving the model structure and completing online optimization. The basic computational unit of a neural network is the neuron, and the formula for calculating the output is as follows: （2） Where X is the input signal of each layer, W is the weight corresponding to the input signal, b is the bias, which can enhance the expressive ability of the neuron, f is the activation function, and C is the output signal of each layer. 4) Train a fully connected neural network, continuously adjusting the network's weights and biases under the stimulation of perceived input, so that the network's output continuously approaches the desired decision result. In the training of the neural network, a non-linear thresholded activation function, the Sigmoid function, is chosen, and its calculation formula is as follows: （3） The formula for calculating the output of the nth hidden layer in a neural network is: （4） Where x m Let ω be the input vector. nm Let b be the weights corresponding to nodes from the m-th layer to the n-th layer, where N is the number of nodes in the n-th layer. n For the bias of the nth layer, Hide n It is the output of the nth layer; The formula for calculating the output layer is: （5） Where C is the output signal of the output layer, ω cn Here, b represents the weights of the nth hidden layer, P is the total number of hidden layers, and b is the weights of the nth hidden layer. c For the bias of the output layer, f c The activation function for the output layer; The residual E is obtained by calculating the mean square error between the output C and the actual label. The calculation formula is as follows: （6） The weights from the output layer to the hidden layer are updated using gradient descent. The impulse of the weights and biases is calculated using the chain rule, and then both are updated separately. The formulas for calculating the impulse of the weights and biases are as follows: （7） （8） in and These are the impulses for the weights and biases, respectively. O is the learning rate, and O is an intermediate function; 5) Finally, update all weights and biases between the hidden layer and the input layer layer by layer. The formula for calculating the impulse of the weights and biases is as follows: （9） （10） in and These are the impulses for the weights and biases, respectively. 6) Evolutionary algorithms encode various parameters of a neural network into chromosomes in the evolutionary algorithm, and then use the evolutionary algorithm to build the network and optimize the parameters. This type of evolutionary neural network uses the weighted topology evolution method, which automatically evolves a neural network structure that meets the requirements based on the perceived input and the human driver's decision output. This includes network weights, network structure and activation functions. 7) The weighted topology evolution method uses an evolutionary neural network based on enhanced topology. First, a complete neural network model including network topology and connection weights is constructed and encoded into node genes and weight genes. 8) Perform mutation and crossover operations, and introduce innovation IDs to record the history of gene crossover and mutation. Mutation operations include four parts: adding a node, adding connections between nodes, adjusting weights, and changing the response of the activation function. Crossover operations follow the principle of "matching gene points are used directly in the child gene, disconnected gene points or excess genes are obtained from the parent gene with high fitness, and two parent genes with the same fitness are randomly inherited." 9) Population dissimilarity distance is defined using disjointed gene points, matched gene points, and excess gene points. The more mismatches there are between two genes, the greater the structural difference between them. The calculation formula is as follows: （11） Where d represents the difference distance, N represents the number of excess gene points in the two genes, M represents the number of disjoint gene points in the two genes, a1, a2 and a3 are coefficients, P is the average of the weight differences of the matching gene points, and L is the length of the longer gene of the two genes. According to this formula, networks with similar structures and weights are assigned to the same group, and intra-group competition occurs during evolution. In this way, populations with new structures do not immediately perish and have sufficient time to adapt to the environment. 10) To avoid a situation where some genes in a group have excessively high fitness to the environment, masking the fitness of other genes with low fitness and reducing the diversity of the entire species, explicit fitness is defined. This means that genes belonging to the same group of species share the same fitness to the environment. The formula for calculating explicit fitness is as follows: （12） Where f i denoted by gene fitness, δ(i,j) represents the difference between two individuals in the group. When δ(i,j) is less than the difference value, sh(δ(i,j)) is 1, and when δ(i,j) is greater than the difference value, sh(δ(i,j)) is 0. n is the total number of individuals in the group. 11) By sharing fitness, each gene is constrained. When there are too many members in a group, its fitness evaluation will be reduced. When a new gene is generated, it will be protected. In this way, if the environmental fitness of an old gene does not improve after multiple generations of evolution, it will be gradually eliminated. 12) The evolutionary algorithm trains the deep learning-based neural network in the neural network-based reasoning module based on the new knowledge reasoning formed between the perceptual input and the human driver's decision output. It adjusts the internal brain-like neuron structure, internal model parameter dimension and organization form to derive the model structure, realize incremental learning of the human-machine hybrid enhanced decision database, and complete the online optimization of the neural network-based reasoning module.

3. The online optimization method for human-machine decision-making logic for highly automated driving according to claim 1, characterized in that: The specific process of step two in the second step is as follows: Human-machine co-driving vehicles are intelligent connected vehicles that connect to the cloud via a network. The vehicle's human-machine hybrid augmented decision-making model is uploaded to the cloud decision model library in real time, and updated in real time from the cloud decision model library. Each human-machine co-driving vehicle performs knowledge reasoning between new perceptual inputs and decision outputs, completing online optimization of case-based reasoning, rule-based reasoning, and neural network-based reasoning modules. The newly updated human-machine hybrid augmented decision-making model is then promptly uploaded to the cloud via the network. Upon receiving the input human-machine hybrid augmented decision-making model, the cloud inputs it into the model judgment module 1 of the cloud decision model library management system. The model judgment module 1 compares the input decision model with the corresponding modules in the current cloud decision model library's case-based, rule-based, and neural network-based reasoning modules. When new driving case libraries, driving rule libraries, or neural networks exist in the model, the model judgment module 1 writes the new parts of the input human-machine hybrid augmented decision model into the corresponding case-based reasoning module library, rule-based reasoning module library, and neural network-based reasoning module library in the cloud decision model library for storage, thereby realizing the updating and optimization of the cloud decision model library. At the same time, the model judgment module 2 of the cloud decision model library management system calls all human-machine hybrid augmented decision models of human-machine co-driving vehicles connected to the cloud and compares the contents of the cloud decision model library with them. When the driving case library, driving rule library, or neural network in the reasoning module of a human-machine hybrid augmented decision model of a human-machine co-driving vehicle is different from that in the cloud decision model library, the reasoning module in the cloud decision model library is downloaded to the human-machine hybrid augmented decision model of that human-machine co-driving vehicle, thereby realizing the updating and online optimization of the human-machine hybrid augmented decision model.

Images