Counterexample-Guided Motion Planner Weight Update
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in generating accurate trajectories that avoid collisions and adhere to desirable characteristics like path length and traffic rules, especially when encountering scenarios not covered during initial training, leading to suboptimal performance in counterexamples.
Innovation Solution
A counterexample-guided update method for motion planners, where the feature weights of a machine learning model trained for one scenario are updated based on those trained for another scenario, optimizing performance across multiple scenarios without retraining, using techniques like Bayesian inference to adjust the probability distribution of feature weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the motion planner is trained for a specific scenario, then it can generate accurate trajectories for that scenario, but it fails to generate correct trajectories for counterexamples (different scenarios)
Solution Approach 1:
The patent applies parameter changes by updating the feature weights of the machine learning model. Instead of retraining the entire model, the system adjusts the weights of features based on counterexample scenarios, allowing the motion planner to adapt to different scenarios while maintaining its original training performance. This resolves the contradiction by enabling scenario coverage expansion without sacrificing trajectory accuracy.
Solution Approach 2:
The patent implements dynamics by making the feature weights dynamic rather than fixed. The system continuously updates feature weights based on encountered scenarios and counterexamples, allowing the motion planner to adapt its behavior dynamically to different driving conditions while maintaining reliability in previously learned scenarios.
2Reliability
If the machine learning model is retrained to handle counterexamples, then performance improves for those scenarios, but the complexity and time required for training increases
Solution Approach 1:
The patent extracts only the essential learning from counterexamples by updating feature weights rather than retraining the entire model. This selective approach captures the necessary adjustments for handling counterexamples while avoiding the complexity of full model retraining, thus improving counterexample performance without proportionally increasing training complexity.
Solution Approach 2:
The patent applies partial action by updating only the feature weights rather than performing complete model retraining. This partial update approach is sufficient to handle counterexamples and improves performance for those scenarios while avoiding the excessive complexity and time consumption associated with full retraining.
3Reliability
If full model retraining is performed to improve counterexample performance, then trajectory accuracy for counterexamples improves, but the time required for training increases
Solution Approach 1:
The patent applies partial action by performing only the necessary weight updates rather than complete model retraining. This partial update approach is sufficient to improve counterexample trajectory accuracy while significantly reducing the time required compared to full retraining, thus resolving the contradiction between accuracy improvement and time loss.
Solution Approach 2:
The patent implements preliminary action by pre-calculating and storing feature weights that can be quickly updated. This allows the system to rapidly adjust to counterexamples using pre-prepared weight structures, improving counterexample accuracy without the time cost of retraining from scratch, thus resolving the time loss contradiction.
Data Source
AI summary
Provided are methods for counterexample-guided update of a motion planner, which may include training a machine learning model to generate a trajectory for a vehicle. The training may generate a first trained machine learning model having a first plurality of feature weights. The machine learning model may be trained based on a counterexample for which the first trained machine learning model fails to generate a correct trajectory for the vehicle. The training may generate a second trained machine learning model having a second plurality of feature weights. A third plurality of feature weights may be determined by updating the first plurality of feature weights based on the second plurality of feature weights. The machine learning model may be updated to generate the trajectory of the vehicle by applying the third plurality of feature weights. Systems and computer program products are also provided.