Search

CN-122020525-A - Automatic driving decision system and method for generating decision tree by using large language model

CN122020525ACN 122020525 ACN122020525 ACN 122020525ACN-122020525-A

Abstract

The invention relates to the field of automatic driving decision, in particular to an automatic driving decision system and method for generating a decision tree by using a large language model. The intelligent driving strategy system comprises an information module and an intelligent agent module, wherein the intelligent agent module comprises a strategy planner, a code generator, a strategy optimizer, a test module and a strategy optimizing module, wherein the strategy planner generates a structured driving strategy text according to a structured data text in information module data, the code generator converts the structured driving strategy text into an executable decision tree code, the strategy optimizing module dynamically adjusts branch weights in the decision tree code to optimize the decision tree code, the test module deploys the decision tree code if the test meets the requirement, and returns to the strategy optimizing module to optimize the decision tree code and adjusts the decision tree code in the strategy planner by adopting the optimized branch weights if the test does not meet the requirement. The invention converts the black box decision in the data-driven method into the decision of generating transparent rules by utilizing the decision tree, thereby solving the problem of the interpretation of the decision.

Inventors

  • WANG PEIHU

Assignees

  • 奇瑞汽车股份有限公司

Dates

Publication Date
20260512
Application Date
20260120

Claims (10)

  1. 1. An automatic driving decision system for generating a decision tree by using a large language model, comprising: The information module is used for converting the input vehicle data and environment data into a structured data text; The intelligent agent module is used for acquiring the structured data text in the information module, generating a decision tree code by adopting the large language model, and generating, executing and optimizing an automatic driving decision according to the decision tree code; The intelligent agent module comprises a strategy planner, a strategy generation module and a strategy generation module, wherein the strategy planner generates a structured driving strategy text according to a structured data text in the information module data; a code generator that converts the structured driving strategy text into executable decision tree code; Dynamically adjusting branch weights in the decision tree codes and optimizing the decision tree codes; And if the test does not meet the requirements, returning to the strategy optimizer to optimize the decision tree code, and adjusting the decision tree code in the strategy planner by adopting the optimized branch weight.
  2. 2. The automated driving decision system utilizing a large language model to generate a decision tree of claim 1, wherein the vehicle data and environmental data input in the information module comprises tracking collected obstacle data, calculated obstacle information, and status data between the vehicle, the obstacle data comprising dynamic obstacle data.
  3. 3. The automatic driving decision system for generating a decision tree by using a large language model according to claim 2, wherein a detection frame, a preset tag and a rule engine are arranged in the information module, the content of the detection frame comprises the obstacle data and the state data, the preset tag is matched with the content of the detection frame, a mark is added for the successfully matched content of the detection frame in a forced mode, and the rule engine is adopted to identify the mark, so that the content of the detection frame is converted into the structured text data.
  4. 4. The automatic driving decision system for generating a decision tree by using a large language model according to claim 1, wherein a prompt word is arranged in the strategy planner, the prompt word is associated with the structured text data and can trigger the large language model to generate the structured driving strategy text, the prompt word comprises the structured text data and can solidify an output format of the structured driving strategy text.
  5. 5. The automated driving decision system utilizing a large language model to generate decision trees of claim 1, wherein the specific flow of the code generator to convert the structured driving strategy text into executable decision tree code comprises: s100, identifying and extracting entity elements in the structured driving strategy text, and converting text information into semantic information required by a simulation test environment; S110, constructing a logic relationship map for the semantic information according to the semantic relationship among the semantic information; S120, generating an object code according to the logic relation graph, and converting the entity element into a control instruction; s130, establishing a multi-level code optimization means to optimize the target code.
  6. 6. The automated driving decision system utilizing a large language model to generate decision trees of claim 5, wherein code boundary processing conditions are provided in the code generator, wherein the code boundary processing conditions include default backup action data in a scenario that is not explicitly defined.
  7. 7. The automated driving decision system utilizing a large language model to generate a decision tree of claim 1, wherein the policy optimizer optimizes the decision tree code, the specific flow comprising: s200, acquiring a complete collision report data packet in the test module, reconstructing the data packet to enable the data packet to comprise a multidimensional time sequence aligned with space time, and simultaneously loading an original decision tree code corresponding to the multidimensional time sequence to generate a complete context analysis text; S210, starting multi-mode root cause analysis based on a three-level diagnosis process based on the complete context analysis text to finish problem positioning; s220, constructing an optimization problem comprising triple constraints, and generating a repair suggestion aiming at the problem; S230, starting a countermeasure verification process, and constructing a specific test scene aiming at each repair suggestion to obtain an optimization strategy; s240, knowledge integration and version management are carried out, the optimized strategy which passes verification is packaged into a knowledge unit and stored in a graph database, and a readability optimized report is generated.
  8. 8. The automated driving decision system for generating decision trees using a large language model as set forth in claim 1 wherein said test module comprises a standard test scenario and a challenge test scenario, wherein said test module loads said decision tree code to be tested, wherein during test operation current environmental observation data is entered into said decision tree code, wherein said decision tree code performs inference calculations based on said current input and outputs control instructions that are sent in real time to a simulation test environment for controlling vehicle execution.
  9. 9. The automated driving decision system using large language model for generating decision tree according to claim 7, wherein the test module comprises the complete collision report packet, the complete collision report packet restores the execution state of the decision tree code, lists key nodes causing collision and input conditions of the key nodes, and extracts complete reasoning path of the decision tree code under collision condition.
  10. 10. An automatic driving decision method for generating a decision tree by using a large language model is applied to an automatic driving decision system according to any one of claims 1-9, and is characterized in that the information module converts the environment data into structured data texts which are understandable by the large language model, the strategy planner module generates the structured driving strategy text according to the structured data texts output by the information module, the code generator module converts the structured driving strategy text output by the strategy planner module into the executable decision tree code, and the strategy optimizer module optimizes the decision tree code.

Description

Automatic driving decision system and method for generating decision tree by using large language model Technical Field The invention relates to the field of automatic driving decision, in particular to an automatic driving decision system and method for generating a decision tree by using a large language model. Background The decision making of automatic driving is related to the safety of driving. Rule-driven automatic driving decision making methods include decision tree methods and expert system methods. The method relies on manual rules and expert knowledge, is transparent to the outside, is easy to modify, has strong interpretability, is difficult to cope with complex dynamic scenes, and cannot be automatically optimized. The real-time performance of the decision method based on the large language model is insufficient, the requirement of the decision frequency of more than 10Hz for automatic driving decision cannot be met, and the large language model has illusion, so that the vehicle is at safety risk. Disclosure of Invention The invention aims to provide an automatic driving decision system and method for generating a decision tree by using a large language model, which combine the decision method based on the large language model with the decision tree method in rule driving, so that the problem of interpretability and instantaneity of the automatic driving decision system can be better solved. In order to achieve the above object, a first aspect of the present invention provides an automatic driving decision system for generating a decision tree by using a large language model, which specifically includes: The information module is used for converting the input vehicle data and environment data into a structured data text; The intelligent agent module is used for acquiring the structured data text in the information module, generating a decision tree code by adopting the large language model, and generating, executing and optimizing an automatic driving decision according to the decision tree code; The intelligent agent module comprises a strategy planner, a strategy generation module and a strategy generation module, wherein the strategy planner generates a structured driving strategy text according to a structured data text in the information module data; a code generator that converts the structured driving strategy text into executable decision tree code; Dynamically adjusting branch weights in the decision tree codes and optimizing the decision tree codes; And if the test does not meet the requirements, returning to the strategy optimizer to optimize the decision tree code, and adjusting the decision tree code in the strategy planner by adopting the optimized branch weight. Optionally, the vehicle data and the environment data input in the information module comprise collected obstacle data, calculated obstacle information and state data between the vehicle, and the obstacle data comprise dynamic obstacle data. Optionally, a detection frame, a preset tag and a rule engine are arranged in the information module, the content of the detection frame comprises the obstacle data and the state data, the preset tag is matched with the content of the detection frame, a tag is added for the successfully matched content of the detection frame in a forced mode, the rule engine is adopted to identify the tag, and the content of the detection frame is converted into the structured text data. Optionally, a prompt word is set in the strategy planner, the prompt word is associated with the structured text data and can trigger the large language model to generate the structured driving strategy text, the prompt word includes the structured text data and can cure an output format of the structured driving strategy text. Optionally, in the code generator, the specific process of converting the structured driving strategy text into executable decision tree code includes: s100, identifying and extracting entity elements in the structured driving strategy text, and converting text information into semantic information required by a simulation test environment; S110, constructing a logic relationship map for the semantic information according to the semantic relationship among the semantic information; S120, generating an object code according to the logic relation graph, and converting the entity element into a control instruction; s130, establishing a multi-level code optimization means to optimize the target code. Optionally, the code generator is provided with code boundary processing conditions, wherein the code boundary processing conditions comprise default backup action data in a scene which is not explicitly defined. Optionally, the policy optimizer optimizes the decision tree code, and the specific flow includes: s200, acquiring a complete collision report data packet in the test module, reconstructing the data packet to enable the data packet to comprise a multidimensional time sequence aligned with spa