Dual-Model Behavior Estimation for Vehicle Merging Decisions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies for controlling mobile object lane changes do not effectively utilize data of a second mobile object, leading to inaccurate estimation of its behavior during merging.
Innovation Solution
An estimation device that utilizes a first trained model and a second trained model to predict the behavior of a second mobile object by comparing actual parameters with distribution information generated from training data collected in scenarios where the second mobile object has and has not taken a predetermined action, using techniques like LSTM to generate predictive distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If passive actor-critic reinforcement learning method is applied to passively collected data, then control policy can be learned, but the behavior of the second mobile object cannot be estimated accurately
Solution Approach 1:
The patent divides the estimation task into two separate trained models: one model estimates behavior when the second mobile object yields, and another model estimates behavior when it does not yield. This segmentation allows each model to specialize in specific behavioral patterns, improving overall estimation accuracy compared to a single unified model.
Solution Approach 2:
The patent employs dynamic modeling by using time-series data to capture the evolving behavior of mobile objects. The trained models process sequential data to adapt to changing traffic conditions and object behaviors, enabling accurate real-time estimation rather than static predictions.
2Ease of manufacture
If data of the second mobile object is not effectively utilized, then existing control policies can be maintained, but estimation accuracy of second mobile object behavior deteriorates
Solution Approach 1:
The patent incorporates feedback mechanisms by training models on historical data that includes actual outcomes of second mobile object behaviors. The models continuously learn from past interactions and use this feedback to improve future predictions, creating a closed-loop system that enhances estimation accuracy over time.
Solution Approach 2:
The patent performs preliminary analysis by training separate models on data specifically labeled for yielding and non-yielding scenarios before actual estimation. This preliminary preparation of specialized models ensures that when estimation is needed, the system can quickly and accurately predict behavior based on the most relevant pre-trained knowledge.
Data Source
AI summary
Provided is an estimation device configured to: acquire prediction parameters, which are distribution information on second parameters, by inputting actual parameters, which are first parameters, into: a first trained model, which is trained to output distribution information on second parameters in response to input of first parameters by using training data and correct data; and a second trained model, which is trained to output distribution information on second parameters in response to input of first parameters by using training data and correct data; and estimate whether or not the second mobile object is to take the predetermined action by comparing the prediction parameters with the actual parameters at a subject time of the prediction parameters.