Autonomous Vehicle Speed Planning With Partitioned Decision Tables
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning-based speed planning for vehicles faces challenges in handling extreme scenarios and requires local adjustments, as the effectiveness of the model is heavily dependent on training data and is difficult to adapt to personalized driving habits and emergency situations.
Innovation Solution
A speed planning method and apparatus that utilizes a machine learning step to obtain a model, followed by a partition decision table approach, allowing for real-time decision-making and local adjustments to control vehicle speed by dividing the input space into partitions and using a partition decision table for quick and localized adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning is used to learn speed planning model from driving data, then the model can adapt to personalized driving habits, but the model effectiveness is heavily dependent on training data quality and cannot handle extreme scenarios
Solution Approach 1:
The patent segments the speed planning system into two parts: a machine learning model for handling common driving scenarios and a rule-based safety constraint layer for ensuring reliability in extreme cases. The learning module processes normal driving patterns while the constraint module independently enforces safety requirements, allowing both adaptability and reliability to coexist without mutual interference.
2Reliability
If manual design and implementation of speed planning model is used to cover all possible scenarios, then comprehensive safety can be ensured, but the model design becomes complex and rare circumstances are likely to be left out
Solution Approach 1:
The patent introduces safety constraints as an intermediary layer between the machine learning model and the final speed control output. This constraint module acts as a mediator that filters and adjusts the learning model's recommendations to ensure they meet safety requirements, thereby achieving comprehensive safety coverage without requiring the entire system to be manually designed.
3Reliability
If the machine learning model is adjusted by adding more training samples to handle particular situations, then the model can be improved, but the training period becomes longer and the final model is not fully predictable
Solution Approach 1:
The patent performs preliminary action by pre-defining safety constraints and rules before the model is deployed. These constraints are established in advance based on safety requirements and regulatory standards, allowing the system to handle particular situations immediately without requiring additional training time. The constraints are built-in from the start rather than added through extended training.
4Adaptability or versatility
If the learning model parameters are adjusted for local situations, then the model can be optimized for specific cases, but the adjustment brings about overall uncontrollable influence
Solution Approach 1:
The patent applies local quality by implementing adjustments at the constraint level rather than modifying the entire learning model. The safety constraints can be locally adjusted for specific situations without affecting the global structure and parameters of the machine learning model, thereby maintaining overall model stability while enabling local optimization.
Data Source
AI summary
A speed planning method and apparatus and a calculating apparatus for automatic driving of a vehicle. The method comprises: using a training sample set to perform machine learning to obtain a machine learning model; partitioning an input space, and obtaining a decision result corresponding to a determined partition based on the obtained machine learning model to form a partition decision table of each partition corresponding to the corresponding decision result; and obtaining each dimensional feature vector of a vehicle while driving in real time as an input feature, determining an input partition to which the input feature belongs, and querying the partition decision table based on the determined partition to obtain the corresponding decision result.