Autonomous Driving Learning Hierarchy With Bidirectional Model Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving systems face challenges in maintaining learning data history and accuracy when new driving environment factors are introduced, leading to the loss of existing learning data and inefficient transfer of learning results across hierarchical layers.
Innovation Solution
The system tiers learning data based on the driving environment, generating root and layer learning data, and performs bidirectional updates between them to maintain data history and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If new driving environment factors are introduced to improve the autonomous driving algorithm, then the adaptability of the system is improved, but the existing learning data is lost and the ability to examine learning history deteriorates
Solution Approach 1:
The learning data is segmented into hierarchical layers (first layer, second layer, third layer, etc.) where each layer corresponds to different levels of driving environment factors. This segmentation allows the system to maintain and examine learning data history at each layer while introducing new environment factors, resolving the contradiction between adaptability and information loss.
2Productivity
If learning data is transferred only from upper to lower layers in a hierarchical structure, then the processing efficiency is improved, but the accuracy of the autonomous driving algorithm deteriorates due to lack of feedback
Solution Approach 1:
The patent implements bidirectional update mechanisms between hierarchical layers. Lower layers can update upper layers when their learning results meet predetermined conditions, creating a feedback loop that improves algorithm accuracy while maintaining processing efficiency. This resolves the contradiction by adding feedback without eliminating the efficient top-down transfer mechanism.
3Ease of manufacture
If all learning data is normalized and processed together, then the manufacturing simplicity is improved, but the ability to examine learning history according to driving environment deteriorates
Solution Approach 1:
The patent segments learning data into hierarchical layers based on driving environment factors, where each layer maintains its specific context while allowing normalized processing within that layer. This segmentation enables both simplified processing within layers and preservation of driving environment context across layers, resolving the contradiction between manufacturing simplicity and information loss.
Data Source
AI summary
Proposed is a method of controlling an autonomous driving system. Root learning data is generated by performing learning for raw data. A plurality of first layer learning data is generated by performing learning, to which driving environment variables of an autonomous vehicle are applied, for the root learning data. The root learning data is updated from the plurality of first layer learning data depending on whether or not an integration condition of the plurality of first layer learning data is met.