Method and apparatus for training intelligent decision model, electronic device and medium

Abstract

The application provides a training method and device of an intelligent decision model, electronic equipment and a medium, and relates to the technical field of artificial intelligence, comprising: obtaining identity information of a driver of a target vehicle, and obtaining driving preference information of the driver based on the identity information of the driver; obtaining historical traffic data of the target vehicle and real-time traffic data and real-time driving data in a driving process; inputting the driving preference information of the driver, the real-time traffic data, the historical traffic data and the real-time driving data into a pre-trained intelligent decision model to obtain at least one recommended driving path; wherein the intelligent decision model adopts a deep network architecture; and determining a target driving path in response to a selection operation of the driver on the recommended driving path. The application improves the explainability and efficiency of path generation and enhances the experience of users.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a training method, apparatus, electronic device, and medium for an intelligent decision-making model. Background Technology

[0002] Artificial intelligence (AI), as a key technology leading a new round of technological revolution and industrial transformation, is profoundly changing people's production and lifestyles. As a new field of social development, AI brings unprecedented opportunities while also posing risks and challenges. Deep learning (Transformer) models, as a highly sought-after AI algorithm, have achieved significant success in various fields such as natural language processing and computer vision. Navigation software commonly utilizes intelligent decision-making models trained on Transformer models for path planning. However, the deep and over-parameterized architecture of Transformer models makes them complex "black box" models, hindering the understanding and interpretability of their internal mechanisms. Low decision-making transparency impacts trust building, and the handling of anomalies is particularly uncontrollable. Therefore, navigation software using intelligent decision-making models trained on Transformer models for path planning suffers from low efficiency and plans that do not meet user needs. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a training method, apparatus, electronic device and medium for an intelligent decision-making model, which improves the interpretability and efficiency of path generation and enhances the user experience.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, this invention provides a path planning method based on an intelligent decision-making model, applied to a navigation system. The intelligent decision-making model adopts a deep network architecture, and the training process includes: obtaining the model architecture parameters of the intelligent decision-making model and initializing the model parameters of the first layer of the intelligent decision-making model; for each layer of the intelligent decision-making model, inputting the output result of the previous layer into the current layer of the intelligent decision-making model for subspace mapping, and constructing the model parameters of the next layer of the intelligent decision-making model based on the input of the current layer, repeating the above process until the last layer of the intelligent decision-making model, to obtain the trained intelligent decision-making model; the method includes: obtaining the identity information of the driver of the target vehicle, and obtaining the driver's driving preference information based on the driver's identity information; obtaining the historical traffic data of the target vehicle and the real-time traffic data and real-time driving data during the driving process; inputting the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data into the pre-trained intelligent decision-making model to obtain at least one recommended driving path; and determining the target driving path in response to the driver's selection operation of the recommended driving path.

[0006] Optionally, the input to the first layer of the intelligent decision-making model is training sample data, which includes at least: vehicle driving data, traffic data, and driver driving preference information; the model architecture parameters include at least: the number of model layers, the update step size, the first model optimization control parameters, the second model optimization control parameters, and the number of model parameter matrices in the first layer; wherein, the model parameters of each layer are represented by a model parameter matrix, each model parameter matrix represents a model subspace, and each model subspace is orthogonal to the others; the training process of the intelligent decision-making model also includes: preprocessing the training sample data to obtain the initial lexical units corresponding to the training sample data, and inputting the initial lexical units into the first layer of the intelligent decision-making model for subspace mapping operation to obtain the output result of the first layer of the intelligent decision-making model.

[0007] Optionally, subspace mapping can be performed according to the following formula:

[0008]

[0009] in, The first intelligent decision-making model represents the... l The output of the layer, The first intelligent decision-making model represents the... l The output of layer -1 The first intelligent decision-making model represents the... l The number of model subspaces in a layer. The first intelligent decision-making model represents the... l Layer model parameters, The function is a normalization function. This indicates the update step size of the intelligent decision-making model.

[0010] Optionally, based on the input of the current layer, the model parameters of the next layer of the intelligent decision-making model are constructed, including: fine-tuning the model parameters of the current layer based on the input of the current layer to obtain the initial model parameters of the next layer of the intelligent decision-making model; optimizing the initial model parameters of the next layer of the intelligent decision-making model based on training sample data; and orthogonalizing the optimized initial model parameters of the next layer of the intelligent decision-making model to obtain the model parameters of the next layer of the intelligent decision-making model.

[0011] Optionally, the model parameters of the current layer are fine-tuned based on the input of the current layer to obtain the initial model parameters of the next layer of the intelligent decision-making model, including: calculating the first process parameters based on the input of the current layer and the model parameters of the current layer; wherein, the first process parameters are: Based on the input of the current layer and the first process parameters, calculate the second process parameters; where the second process parameters are: ; Calculate the normal form value of each column of the second process parameters, and determine the initial model parameters of the next layer of the intelligent decision-making model based on the normal form value of each column of the second process parameters.

[0012] Optionally, based on the normal form values ​​of each column of the second process parameters, the initial model parameters for the next layer of the intelligent decision-making model are determined, including: obtaining non-zero vectors from the current layer's model parameters in descending order of normal form values; wherein the preset column is the number of columns of the current layer's model parameters plus one; if the normal form values ​​corresponding to the non-zero vectors in the preset column are all greater than the first model optimization control parameters, then the number of columns of the next layer's model parameters is set equal to the preset column; otherwise, the normal form value corresponding to the non-zero vector at the preset value position is obtained in descending order of normal form values; wherein the preset value is the number of columns of the current layer's model parameters; if the normal form value corresponding to the non-zero vector at the preset value position is less than the second model... To optimize the control parameters, the number of columns in the model parameters of the next layer is equal to the number of columns in the model parameters of the current layer minus one; otherwise, the number of columns in the model parameters of the next layer is equal to the number of columns in the model parameters of the current layer. Following the order of normality values ​​from largest to smallest, non-zero vectors representing the number of columns in the model parameters of the next layer are read from the model parameters of the current layer to form the initial model parameters of the next layer. If the model parameters of the current layer are a zero matrix, the number of model subspaces in the next layer is equal to the number of model subspaces in the current layer; otherwise, the number of model subspaces in the next layer is equal to the number of model subspaces in the current layer plus one, and the non-zero vectors in the model parameters of the current layer are incorporated into the initial model parameters of the next layer.

[0013] Optionally, the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data are input into a pre-trained intelligent decision-making model to obtain at least one recommended driving route, including: determining the driver's travel intention based on the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data; generating multiple candidate driving routes based on the travel intention; evaluating the candidate driving routes; and obtaining at least one recommended driving route.

[0014] Optionally, multiple candidate driving routes are generated based on travel intent, and the candidate driving routes are evaluated to obtain at least one recommended driving route. This includes: generating an initial driving route using an improved graph search algorithm, and generating multiple candidate driving routes based on travel intent and the initial driving route; for each candidate driving route, extracting feature information; wherein the feature information includes at least: congestion index, predicted travel time, distance, number of traffic lights, driver preference matching degree, and weather information for each road segment; calculating a comprehensive score for each candidate driving route based on the feature information of the candidate driving routes, and determining at least one recommended driving route based on the comprehensive score.

[0015] Secondly, this invention provides a path planning device based on an intelligent decision-making model, applied to a navigation system. The intelligent decision-making model adopts a deep network architecture, and the training process includes: acquiring the model architecture parameters of the intelligent decision-making model and initializing the model parameters of the first layer of the intelligent decision-making model; for each layer of the intelligent decision-making model, inputting the output result of the previous layer into the current layer of the intelligent decision-making model for subspace mapping, and constructing the model parameters of the next layer of the intelligent decision-making model based on the input of the current layer, repeating the above process until the last layer of the intelligent decision-making model, to obtain the trained intelligent decision-making model; the device includes: a driving preference information acquisition module, used to acquire the identity information of the driver of the target vehicle, and based on the driver's identity information, acquire the driver's driving preference information; a driving data acquisition module, used to acquire the historical traffic data of the target vehicle and the real-time traffic data and real-time driving data during the driving process; a path planning module, used to input the driver's driving preference information, real-time traffic data, historical traffic data and real-time driving data into the pre-trained intelligent decision-making model to obtain at least one recommended driving path; and a path determination module, used to determine the target driving path in response to the driver's selection operation of the recommended driving path.

[0016] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the steps of the method provided in any of the first aspects above.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the method provided in any of the first aspects above.

[0018] This invention brings the following beneficial effects:

[0019] The training method, apparatus, electronic device, and medium for the intelligent decision-making model provided by this invention are applied to a navigation system. First, the identity information of the driver of the target vehicle is obtained, and based on this identity information, the driver's driving preference information is obtained. Then, historical traffic data of the target vehicle, as well as real-time traffic data and real-time driving data during the driving process, are obtained. Next, the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data are input into a pre-trained intelligent decision-making model to obtain at least one recommended driving route. The intelligent decision-making model employs a deep network architecture. Finally, in response to the driver's selection of the recommended driving route, the target driving route is determined. In this method, by combining information such as the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data, and using an intelligent decision-making model constructed with a deep network architecture, at least one recommended driving route is generated, improving the interpretability and efficiency of route generation and enhancing the user experience.

[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a path planning method based on an intelligent decision-making model, provided as an embodiment of the present invention;

[0024] Figure 2 A flowchart of a Transformer construction method provided in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram illustrating the construction and parsing of a Transformer according to an embodiment of the present invention;

[0026] Figure 4 A schematic diagram of the structure of a path planning device based on an intelligent decision-making model provided in an embodiment of the present invention;

[0027] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Currently, navigation software utilizes intelligent decision-making models trained on Transformer models for route planning, a common application scenario for Transformer models. However, the deep and over-parameterized architecture of the Transformer model makes it a complex "black box," hindering the understanding and interpretability of its internal mechanisms. Low decision-making transparency impacts trust building, and the handling of anomalies is particularly uncontrollable. Therefore, navigation software using intelligent decision-making models trained on Transformer models for route planning suffers from low efficiency and routes that do not meet user needs.

[0030] Based on this, the present invention provides a training method, apparatus, electronic device and medium for an intelligent decision-making model, which improves the interpretability and efficiency of path generation and enhances the user experience.

[0031] To facilitate understanding of this embodiment, a path planning method based on an intelligent decision-making model disclosed in this invention will first be described in detail. This method can be executed by electronic devices, such as smartphones, computers, and tablets. See also Figure 1 The flowchart shown illustrates a path planning method based on an intelligent decision-making model, indicating that the method mainly includes the following steps S101 to S104:

[0032] Step S101: Obtain the identity information of the driver of the target vehicle, and based on the driver's identity information, obtain the driver's driving preference information.

[0033] In one implementation, the driver's facial image can be captured by a camera inside the target vehicle, and the captured facial image can be compared with pre-stored driver image information to determine the driver's identity. Alternatively, the driver's identity can be determined using the current login information of the target vehicle's navigation system. Furthermore, based on the driver's identity information, the driver's driving preference information can be obtained, which includes at least: frequent use of highways, avoidance of toll roads, time sensitivity or greater focus on fuel consumption / distance, frequent departure / arrival times, frequently visited locations, and historical selection behavior.

[0034] Step S102: Obtain historical traffic data of the target vehicle, as well as real-time traffic data and real-time driving data during the driving process.

[0035] In one implementation, historical traffic data of the target vehicle can be read from the navigation system, including at least: fixed congestion on a main road during weekday morning rush hour, afternoon congestion around scenic spots on weekends, and decreased traffic efficiency on certain road sections under specific weather conditions (such as rain or snow); real-time traffic data can be obtained from traffic sensors, floating car data, traffic management platforms, etc., including at least: road network topology (nodes, road segments, connection relationships), real-time vehicle speed of each road segment, congestion level, accident or construction information, traffic light timing, temporary road closures, and other dynamic events; and real-time driving data can be obtained, including: origin, destination, departure time, weather, holidays, etc.

[0036] Step S103: Input the driver's driving preference information, real-time traffic data, historical traffic data and real-time driving data into the pre-trained intelligent decision-making model to obtain at least one recommended driving route.

[0037] In one implementation, the intelligent decision-making model employs a deep network architecture. Driver preference information, real-time traffic data, historical traffic data, and real-time driving data are input into a pre-trained intelligent decision-making model, which can then generate at least one recommended driving route based on the input data.

[0038] Step S104: In response to the driver's selection of the recommended driving route, determine the target driving route.

[0039] In one implementation, the driver can select one of the recommended driving routes output by the intelligent decision-making model as the target driving route.

[0040] In addition, the driver's actual route selection, whether navigation is canceled midway, and the deviation between the actual arrival time and the prediction will be recorded as feedback signals. The system can use the feedback signals to fine-tune the intelligent decision-making model to improve the overall prediction accuracy.

[0041] The path planning method based on the intelligent decision-making model provided in this embodiment of the invention combines information such as the driver's driving preferences, real-time traffic data, historical traffic data, and real-time driving data. By using an intelligent decision-making model constructed with a deep network architecture, it generates at least one recommended driving route, which improves the interpretability and efficiency of path generation and enhances the user experience.

[0042] In one implementation, step S103, which involves inputting the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data into a pre-trained intelligent decision-making model to obtain at least one recommended driving route, can be achieved in ways including but not limited to the following steps, mainly including steps 1 to 2:

[0043] Step 1: Determine the driver's travel intention based on the driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data.

[0044] In practice, contextual factors such as driver preferences, real-time traffic data, historical traffic data, and real-time driving data are uniformly encoded into a comprehensive scenario description vector, which serves as input to the intelligent decision-making model. The intelligent decision-making model deeply understands the input context to infer the driver's travel intentions. For example, if a driver is heading home from work during rush hour, even without explicit explanation, the model may assume they want to get home as quickly as possible; if a driver repeatedly chooses scenic detours on weekends, it may be identified as a leisure travel scenario.

[0045] Step 2: Generate multiple candidate driving routes based on travel intentions, evaluate the candidate driving routes, and obtain at least one recommended driving route.

[0046] In practice, firstly, an improved graph search algorithm is used to generate an initial driving route, and based on the travel intention and the initial driving route, multiple candidate driving routes are generated. Then, for each candidate driving route, feature information is extracted. The feature information includes at least: congestion index of each road segment, predicted travel time, distance, number of traffic lights, driver preference matching degree, and weather information. Finally, based on the feature information of the candidate driving routes, a comprehensive score is calculated for each candidate driving route, and based on the comprehensive score, at least one recommended driving route is determined.

[0047] Specifically, firstly, an improved graph search algorithm (such as Dijkstra's algorithm) is used to quickly generate a batch of initial driving routes in the road network topology. These initial driving routes meet basic constraints, such as connectivity from the origin to the destination, reasonable length or time, and conformity to the user's explicit preferences. Then, combined with the driver's travel intentions, multiple candidate driving routes are generated. For example, if the driver values ​​speed, the focus is on main roads and high-traffic-efficiency road sections, generating multiple variations under different highway / expressway combinations. If the driver prefers familiar routes or frequently visited areas, priority is given to searching areas covered by the user's historical trajectory, or routes passing through frequently used locations are given a higher generation probability. If it is a special vehicle (such as a truck or electric vehicle), restricted sections, steep slopes, and areas without charging stations are excluded, and candidates are generated in the compliant subgraph.

[0048] To avoid recommending multiple highly overlapping paths, this embodiment of the invention can also employ topology separation to force candidate paths to make different choices at key nodes (such as overpasses and regional boundaries); or perform similarity clustering on the initially generated large number of paths, retaining only the best representative for each class.

[0049] Furthermore, for each candidate driving route, corresponding feature information is extracted. This feature information includes, but is not limited to: current speed of each road segment, congestion index, predicted travel time, impact of emergencies such as accidents, construction, and road closures, total distance, number of traffic lights, number of turns, highway / urban road ratio, speed limit changes, road condition stability (some shorter roads are prone to congestion), driver preference matching degree (e.g., whether it includes frequently used road segments, whether it matches the user's historical selection patterns, etc.), departure time, weather conditions, holidays, etc. Finally, each candidate driving route is comprehensively scored based on the feature information, and at least one recommended driving route is determined based on the comprehensive score.

[0050] In one implementation, when calculating a comprehensive score for each candidate driving route based on its feature information, and determining at least one recommended driving route based on the comprehensive score, the following methods can be used: First, determine the weight coefficients of the feature information based on the travel intention and / or the road segment information of the candidate driving route; then, calculate the comprehensive score for each candidate driving route based on its feature information and the corresponding weight coefficients; finally, rank the candidate driving routes based on the comprehensive score, and determine at least one recommended driving route based on the ranking result.

[0051] In practical implementation, the model can automatically identify the bottleneck sections in the route that have the greatest impact on the overall experience (such as a perpetually congested overpass) and assign them higher weight coefficients. It can also dynamically adjust the importance of feature information in different scenarios based on travel intentions. For example, in emergency medical transport scenarios, the fastest arrival has a very high weight coefficient; in nighttime leisure travel, the smoothest and least disturbed route has a higher weight coefficient. In addition, the model can also evaluate the reliability of the prediction. If a route depends on a rural road with sparse data, its recommendation priority will be reduced even if the apparent time is short.

[0052] Furthermore, the model can calculate a comprehensive score for each candidate driving route based on its feature information and corresponding weight coefficients. Then, it can sort the candidate driving routes according to the comprehensive score and select the Top-K routes (usually 2–3) as recommended driving routes.

[0053] For ease of understanding, this embodiment of the invention also provides a training process for the intelligent decision-making model, which mainly includes the following steps 1 to 3:

[0054] Step 1: Obtain the model architecture parameters of the intelligent decision-making model and initialize the model parameters of the first layer of the intelligent decision-making model.

[0055] In one implementation, the model architecture parameters include at least: the number of model layers. Update step size First model optimization control parameters Second model optimizes control parameters The number of model parameter matrices in the first layer In this model, the model parameters of each layer are represented by a model parameter matrix, and each model parameter matrix represents a model subspace. The model subspaces are orthogonal to each other.

[0056] Specifically, update step size Indicates the magnitude of updates and iterations between model layers; first model optimization control parameters Second model optimizes control parameters It indicates the degree of model adjustment and optimization; the parameters of each layer of the model are a set of orthogonal bases, represented by multiple matrices, each matrix representing a model subspace.

[0057] Furthermore, based on the above model architecture parameters, the model parameters for the first layer of the intelligent decision-making model are: Model parameter matrix , That is, the model in the first layer Each model subspace. Satisfying mutual orthogonality and , It is a positive integer. In this embodiment of the invention, the parameters of the first layer of the intelligent decision-making model can be randomly generated or obtained through pre-training.

[0058] Step 2: Preprocess the training sample data to obtain the initial word units corresponding to the training sample data, and input the initial word units into the first layer of the intelligent decision model for subspace mapping operation to obtain the output result of the first layer of the intelligent decision model.

[0059] In one implementation, the input to the first layer of the intelligent decision-making model is training sample data, which includes at least: driver's driving preference information, real-time traffic data, historical traffic data, real-time driving data, and driving routes. Assume the input training sample data is: ,in, for token for The number of columns. For Preprocessing, such as position encoding, word embedding, etc., will... Convert to initial lexicon The dimension of each word element is changed from Become .

[0060] Specifically, subspace mapping operations can be performed according to the following formula:

[0061] (1)

[0062] in, The first intelligent decision-making model represents the... l The output of the layer, The first intelligent decision-making model represents the... l The output of layer -1 The first intelligent decision-making model represents the... l The number of model subspaces in a layer. . The first intelligent decision-making model represents the... l Layer model parameters, The function is a normalization function. This indicates the update step size of the intelligent decision-making model.

[0063] As shown in the formula above, the intelligent decision-making model only includes subspace mapping operations and jump connections. Each computation completes a subspace mapping, increasing the mapped space... The weight of its own components, while decreasing Weighting of noise, for After parsing and noise reduction, the result is .

[0064] Step 3: For each layer of the intelligent decision model, input the output of the previous layer into the current layer of the intelligent decision model for subspace mapping, and construct the model parameters of the next layer of the intelligent decision model based on the input of the current layer. Repeat the above process until the last layer of the intelligent decision model to obtain the trained intelligent decision model.

[0065] In one implementation, the output of the previous layer of the model is used as the input of the current layer. After subspace mapping of the input of the model layer by layer, the model parameters of the next layer are constructed until the last layer of the intelligent decision model, thereby obtaining the model parameters of each layer and thus obtaining the trained intelligent decision model.

[0066] The training method for the intelligent decision-making model provided in this embodiment of the invention first initializes the parameters of the first layer of the model. Then, it performs a subspace mapping operation on the input of each layer of the intelligent decision-making model and constructs the model parameters for the next layer after mapping the input. When the mapping reaches the last layer, the intelligent decision-making model is formed. The intelligent decision-making model trained by the above method has a simple structure, strong input-output interpretability, and ensures the performance of the model while improving its reliability and controllability.

[0067] In one implementation, for step 3 above, i.e., when constructing the model parameters for the next layer of the intelligent decision-making model based on the input of the current layer, the following methods can be used, including but not limited to:

[0068] First, the model parameters of the current layer are fine-tuned based on the input of the current layer to obtain the initial model parameters of the next layer of the intelligent decision-making model.

[0069] In practical implementation, the first process parameters are calculated based on the input and model parameters of the current layer; the first process parameters are: Then, based on the input of the current layer and the first process parameters, the second process parameters are calculated; wherein, the second process parameters are: Finally, the normal form value of each column of the second process parameters is calculated, and based on the normal form value of each column of the second process parameters, the initial model parameters for the next layer of the intelligent decision-making model are determined.

[0070] Specifically, in the first stage of model building... l Model parameters of +1 layer When, the current layer is the th l Layer, according to the model l Layer input For the first l Layer model parameters Fine-tuning is performed to obtain new parameters, namely the first... l +1 layer model parameters.

[0071] The model's first l The model parameters of the layer are represented as follows .make For the model number l The model parameters of the layers are connected. This allows for the calculation of the first process parameters: For all All of them are operated as follows to build Calculate the parameters for the second process: Calculate the normal form value of each column of the second process parameters. Based on the normal form values ​​of each column of the second process parameters, the initial model parameters for the next layer of the intelligent decision-making model are determined.

[0072] In practical implementation, firstly, following the order of normality values ​​from largest to smallest, obtain the non-zero vectors of a preset column from the model parameters of the current layer; where the preset column is the column number of the model parameters of the current layer plus one. Specifically, because... Then, according to the order of normal form values ​​from largest to smallest, take... middle The highest value A non-zero vector of columns (i.e., preset columns), where, For the first l The number of columns in the model parameters of the layer.

[0073] If the normal form values ​​corresponding to the non-zero vectors in the preset column are all greater than the first model optimization control parameter, then the number of columns of the model parameters in the next layer is set to equal the preset column. Otherwise, the normal form values ​​corresponding to the non-zero vectors at the preset position are obtained in descending order of normal form values; where the preset value is the number of columns of the model parameters in the current layer. Specifically, if The column non-zero vector corresponding to The norm values ​​are all greater than the first model's optimized control parameters. Then, set the number of columns in the model parameters of the next layer to equal the preset number of columns, i.e. Otherwise, obtain the value in descending order of normality value. The normal form value corresponding to a non-zero vector (i.e., a preset value).

[0074] If the normal form value corresponding to the non-zero vector at the preset value position is less than the second model optimization control parameter, then the number of columns of the model parameters in the next layer is equal to the number of columns of the model parameters in the current layer minus one; otherwise, the number of columns of the model parameters in the next layer is equal to the number of columns of the model parameters in the current layer. Specifically, if the vector at the preset value position is... The normal form value corresponding to the non-zero vector of a bit is less than the second model optimization control parameter. Then, let the number of columns of the model parameters in the next layer be equal to the number of columns of the model parameters in the current layer minus one, that is... , for other cases, let the number of columns of the model parameters of the next layer be equal to the number of columns of the model parameters of the current layer, that is .

[0075] Furthermore, in the order from the largest to the smallest of the norm values, read the non-zero vectors of the number of columns of the model parameters of the next layer from the model parameters of the current layer to form the initial model parameters of the next layer. Specifically, take the model parameters of the current layer in the columns with the highest values to form the initial model parameters of the next layer , and the vectors incorporated into

[0076] become zero vectors.

[0077] Specifically, when is built, if is a zero matrix, then let , otherwise , and incorporate the non-zero vectors in into to complete the construction of the new model parameters .

[0078] Then, optimize the initial model parameters of the next layer of the intelligent decision-making model based on the training sample data. Specifically, can be optimized by training methods such as backpropagation according to the training sample data.

[0079] Finally, perform orthogonalization on the optimized initial model parameters of the next layer of the intelligent decision-making model to obtain the model parameters of the next layer of the intelligent decision-making model. Specifically, perform orthogonalization on to satisfy being orthogonal to each other and , and form the model parameters of the l +1 layer . When l <L, let l = l +1, and continue to build the model parameters. Otherwise, the output of the last layer of the model is the exact analysis of , and the model parameters ​The operation method that satisfies formula (1) constitutes the L-layer white-box Transformer model, which is the final intelligent decision model.

[0080] For ease of understanding, this embodiment of the invention also provides a flowchart of a Transformer constructor method, see [link / reference]. Figure 2 As shown, the main processes include the following:

[0081] (1) Preprocess the input. Specifically, assume the input is: ,in for token for The number of columns. For Preprocessing, such as position encoding, word embedding, etc., will... Convert to initial lexicon The dimension of each word element is changed from Become .

[0082] (2) Set model architecture parameters. Specifically, the model architecture parameters should include at least the number of model layers. Update step size First model optimization control parameters Second model optimizes control parameters The number of model parameter matrices in the first layer .

[0083] (3) Initialize the parameters of the first layer of the model. Specifically, the parameters of the first layer of the model are: Model parameter matrix , That is, the model in the first layer Each model subspace.

[0084] (4) Calculation model l Layer output. Specifically, the model's output is calculated according to the aforementioned formula (1). l Layer output.

[0085] (5) Constructing the model l Add 1 more parameters to the model and let l = l + 1. Specifically, according to the model's... l Layer input Model parameters for layer l Fine-tuning was performed to obtain the new parameters. The specific process can be found in the aforementioned embodiments, and will not be repeated here.

[0086] (6) If the model reaches the last layer, the model is completed and output; otherwise, return (4).

[0087] Furthermore, this invention also provides a schematic diagram of Transformer construction and input parsing, see [link to schematic diagram]. Figure 3 As shown. For the input Preprocessing is performed to obtain the input of the first layer of the model. Initialize the parameters of the first layer of the model. , thus obtaining the first layer output After that, by and Wait to construct the second layer of parameters , thus obtaining the first layer output And so on, from and Construct the first Layer parameters until obtained .

[0088] The method provided in this invention simplifies the model architecture to include only subspace mapping operations and skip connections. The model parameters for each layer are composed of a set of orthogonal bases. While the parameters differ between layers, they all employ the same subspace denoising operation. This operation, on the one hand, parses the important components contained in the input, and on the other hand, amplifies the difference between the important components and the noise. This allows each layer of the model to continuously parse the input while simultaneously denoising the noise. Finally, the model output is a precisely parsed and denoised input, facilitating the execution of subsequent downstream tasks. This method enhances the safety of the Transformer by constructing a simple, highly interpretable Transformer architecture without sacrificing performance.

[0089] In addition to the path planning method based on the intelligent decision model provided in the foregoing embodiments, this invention also provides a path planning device based on the intelligent decision model, which is applied to a navigation system. The intelligent decision model adopts a deep network architecture, and the training process includes: obtaining the model architecture parameters of the intelligent decision model and initializing the model parameters of the first layer of the intelligent decision model; for each layer of the intelligent decision model, inputting the output result of the previous layer into the current layer of the intelligent decision model for subspace mapping, and constructing the model parameters of the next layer of the intelligent decision model based on the input of the current layer, repeating the above process until the last layer of the intelligent decision model, to obtain the trained intelligent decision model.

[0090] See Figure 4 The diagram shown illustrates the structure of a path planning device based on an intelligent decision-making model, indicating that the device mainly comprises the following parts:

[0091] The driving preference information acquisition module 401 is used to acquire the identity information of the driver of the target vehicle and, based on the driver's identity information, acquire the driver's driving preference information.

[0092] The driving data acquisition module 402 is used to acquire historical traffic data of the target vehicle as well as real-time traffic data and real-time driving data during the driving process.

[0093] The route planning module 403 is used to input the driver's driving preference information, real-time traffic data, historical traffic data and real-time driving data into a pre-trained intelligent decision-making model to obtain at least one recommended driving route; wherein, the intelligent decision-making model adopts a deep network architecture.

[0094] The route determination module 404 is used to determine the target driving route in response to the driver's selection of the recommended driving route.

[0095] The path planning device based on the intelligent decision-making model provided in this embodiment of the invention combines information such as the driver's driving preferences, real-time traffic data, historical traffic data, and real-time driving data. By using an intelligent decision-making model constructed with a deep network architecture, it generates at least one recommended driving route, which improves the interpretability and efficiency of path generation and enhances the user experience.

[0096] It should be noted that the device provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment. The specific numerical values ​​provided in this embodiment are merely exemplary and are not intended to limit the scope of the invention.

[0097] This invention also provides an electronic device, specifically, the electronic device includes a processor and a storage device; the storage device stores a computer program, and the computer program, when run by the processor, executes the method described in any of the above embodiments.

[0098] Figure 5 The present invention provides a schematic diagram of the structure of an electronic device 100, which includes a processor 50, a memory 51, a bus 52 and a communication interface 53. The processor 50, the communication interface 53 and the memory 51 are connected through the bus 52. The processor 50 is used to execute executable modules, such as computer programs, stored in the memory 51.

[0099] The memory 51 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 53 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0100] Bus 52 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0101] The memory 51 is used to store programs. After receiving an execution instruction, the processor 50 executes the programs. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 50 or implemented by the processor 50.

[0102] Processor 50 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 50 or by instructions in software form. Processor 50 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 51. The processor 50 reads the information in memory 51 and, in conjunction with its hardware, completes the steps of the above method.

[0103] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0104] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A path planning method based on an intelligent decision-making model, characterized in that, Applied to navigation systems, the intelligent decision-making model employs a deep network architecture. The training process includes: obtaining the model architecture parameters of the intelligent decision-making model and initializing the model parameters of the first layer; for each layer of the intelligent decision-making model, inputting the output of the previous layer into the current layer for subspace mapping, and constructing the model parameters of the next layer based on the input of the current layer; repeating the above process until the last layer of the intelligent decision-making model, thus obtaining the trained intelligent decision-making model; the method includes: Obtain the identity information of the driver of the target vehicle, and based on the driver's identity information, obtain the driver's driving preference information; Acquire historical traffic data of the target vehicle, as well as real-time traffic data and real-time driving data during the driving process; The driver's driving preference information, the real-time traffic data, the historical traffic data, and the real-time driving data are input into a pre-trained intelligent decision-making model to obtain at least one recommended driving route; In response to the driver's selection of the recommended driving route, a target driving route is determined.

2. The method according to claim 1, characterized in that, The input to the first layer of the intelligent decision-making model is training sample data, which includes at least: vehicle driving data, traffic data, and driver driving preference information; the model architecture parameters include at least: the number of model layers, update step size, first model optimization control parameters, second model optimization control parameters, and the number of model parameter matrices in the first layer; wherein, the model parameters of each layer are represented by a model parameter matrix, each model parameter matrix represents a model subspace, and each model subspace is orthogonal to the others; the training process of the intelligent decision-making model further includes: The training sample data is preprocessed to obtain the initial word units corresponding to the training sample data, and the initial word units are input into the first layer of the intelligent decision model for subspace mapping operation to obtain the output result of the first layer of the intelligent decision model.

3. The method according to claim 2, characterized in that, Subspace mapping operations are performed according to the following formula: in, The first term of the intelligent decision-making model is represented by... l The output of the layer, The first term of the intelligent decision-making model is represented by... l The output of layer -1 The first term of the intelligent decision-making model is represented by... l The number of model subspaces in a layer. The first term of the intelligent decision-making model is represented by... l Layer model parameters, The function is a normalization function. This indicates the update step size of the intelligent decision-making model.

4. The method according to claim 3, characterized in that, Based on the input of the current layer, the model parameters for the next layer of the intelligent decision-making model are constructed, including: Based on the input of the current layer, the model parameters of the current layer are fine-tuned to obtain the initial model parameters of the next layer of the intelligent decision-making model; The initial model parameters of the next layer of the intelligent decision-making model are optimized based on the training sample data. The initial model parameters of the next layer of the optimized intelligent decision-making model are orthogonalized to obtain the model parameters of the next layer of the intelligent decision-making model.

5. The method according to claim 4, characterized in that, Based on the input of the current layer, the model parameters of the current layer are fine-tuned to obtain the initial model parameters of the next layer of the intelligent decision-making model, including: Based on the input of the current layer and the model parameters of the current layer, the first process parameters are calculated; wherein, the first process parameters are: ; Based on the input of the current layer and the first process parameter, the second process parameter is calculated; wherein the second process parameter is: ; Calculate the normal form value of each column of the second process parameters, and determine the initial model parameters for the next layer of the intelligent decision-making model based on the normal form value of each column of the second process parameters.

6. The method according to claim 5, characterized in that, Based on the normal form value of each column of the second process parameters, the initial model parameters for the next layer of the intelligent decision-making model are determined, including: According to the order of the norm values ​​from largest to smallest, obtain the non-zero vector of a preset column from the model parameters of the current layer; wherein, the preset column is the number of columns of the model parameters of the current layer plus one; If the normal form values ​​corresponding to the non-zero vectors in the preset column are all greater than the first model optimization control parameter, then the number of columns of the model parameters in the next layer is set to be equal to the preset column; otherwise, the normal form values ​​corresponding to the non-zero vectors at the preset position are obtained in descending order of the normal form values; wherein, the preset value is the number of columns of the model parameters in the current layer. If the normal form value corresponding to the non-zero vector at the preset value position is less than the second model optimization control parameter, then the number of columns of the model parameters in the next layer is equal to the number of columns of the model parameters in the current layer minus one; otherwise, the number of columns of the model parameters in the next layer is equal to the number of columns of the model parameters in the current layer. According to the order of the norm values ​​from largest to smallest, the non-zero vector of the column number of the model parameters of the next layer is read from the model parameters of the current layer to form the initial model parameters of the next layer; If the model parameters of the current layer are zero matrices, then the number of model subspaces in the next layer is equal to the number of model subspaces in the current layer; otherwise, the number of model subspaces in the next layer is equal to the number of model subspaces in the current layer plus one, and the non-zero vectors in the model parameters of the current layer are incorporated into the initial model parameters of the next layer.

7. The method according to claim 1, characterized in that, The driver's driving preference information, real-time traffic data, historical traffic data, and real-time driving data are input into a pre-trained intelligent decision-making model to obtain at least one recommended driving route, including: Based on the driver's driving preference information, the real-time traffic data, the historical traffic data, and the real-time driving data, the driver's travel intention is determined; Multiple candidate driving routes are generated based on the travel intention, and the candidate driving routes are evaluated to obtain at least one recommended driving route.

8. The method according to claim 7, characterized in that, Multiple candidate driving routes are generated based on the travel intention, and the candidate driving routes are evaluated to obtain at least one recommended driving route, including: An improved graph search algorithm is used to generate an initial driving route, and multiple candidate driving routes are generated based on the travel intention and the initial driving route. For each candidate driving route, feature information of the candidate driving route is extracted; wherein, the feature information includes at least: congestion index of each road segment, predicted travel time, distance, number of traffic lights, driver preference matching degree, and weather information; Based on the feature information of the candidate driving routes, a comprehensive score is calculated for each candidate driving route, and based on the comprehensive score, at least one recommended driving route is determined.

9. A path planning device based on an intelligent decision-making model, characterized in that, Applied to navigation systems, the intelligent decision-making model employs a deep network architecture. The training process includes: acquiring the model architecture parameters of the intelligent decision-making model and initializing the model parameters of the first layer; for each layer of the intelligent decision-making model, inputting the output of the previous layer into the current layer for subspace mapping, and constructing the model parameters of the next layer based on the input of the current layer; repeating the above process until the last layer of the intelligent decision-making model, thus obtaining the trained intelligent decision-making model; the device includes: The driving preference information acquisition module is used to acquire the identity information of the driver of the target vehicle, and based on the driver's identity information, acquire the driver's driving preference information. The driving data acquisition module is used to acquire the target vehicle's historical traffic data, as well as real-time traffic data and real-time driving data during the driving process; The route planning module is used to input the driver's driving preference information, the real-time traffic data, the historical traffic data, and the real-time driving data into a pre-trained intelligent decision-making model to obtain at least one recommended driving route. The route determination module is used to determine the target driving route in response to the driver's selection of the recommended driving route.

10. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the steps of the method according to any one of claims 1 to 8.

Images