Bicycle Control Model Interpolation for Rider-Adaptive Automation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing human-powered vehicle control systems require extensive training time and struggle to adapt to untrained situations, especially when dealing with individual rider variations and different vehicle types or environments.
Innovation Solution
A human-powered vehicle control device that utilizes a first learning model interpolated by a second learning model, trained on different vehicles or riders, to reduce training time and enable automatic control in unlearned situations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning is used to train a model for automatic control of human-powered vehicle devices, then automatic control can be optimized for individual riders, but the training time becomes excessively long
Solution Approach 1:
The patent applies preliminary action by pre-training learning models offline before the vehicle is actually used. Multiple learning models are trained in advance using training data from various riders and vehicles, and these pre-trained models are stored for later use. This allows the system to have ready-to-use models without requiring lengthy training during actual operation, thus resolving the contradiction between achieving accurate automatic control and avoiding excessive training time.
2Reliability
If a learning model is trained individually for each rider and vehicle, then control can be optimized, but the system cannot handle untrained situations
Solution Approach 1:
The patent implements universality by creating a plurality of learning models that can be universally applied across different riders and vehicle types. Instead of training a single specialized model for each individual case, the system trains multiple models with different characteristics and stores them all. When a situation arises, the system can select from among these pre-trained models to handle various scenarios including untrained situations, thus achieving both optimization and adaptability.
Solution Approach 2:
The patent applies copying by creating multiple copies of learning models trained on different datasets. Rather than having one original model, the system generates multiple model copies with varying training backgrounds. These copied models can then be selected based on the current situation, allowing the system to handle both familiar and novel situations effectively without requiring real-time training.
3Measurement precision
If learning is performed based on individual rider and vehicle characteristics, then personalized control is achieved, but situations not covered in training data cannot be handled
Solution Approach 1:
The patent applies parameter changes by varying the training parameters and conditions when creating multiple learning models. Different models are trained with different hyperparameters, learning rates, network architectures, or training data subsets. This creates a diverse set of models with different characteristics that can handle various situations. When deployed, the system can select the model whose parameters best match the current situation, thereby achieving both personalized accuracy and broad situational coverage.
Data Source
AI summary
A human-powered vehicle control device includes at least one sensor, a memory, a controller and an interpolation processor. The at least one sensor is configured to acquire input information related to traveling of a human-powered vehicle. The memory is configured to store a first learning model trained so as to output output information related to control of a device mounted on the human-powered vehicle based on the input information acquired. The controller is configured to control the device by control data decided based on output information obtained by inputting the input information to the first learning model. The interpolation processor is configured to execute processing of interpolating the first learning model in the memory using a second learning model trained with input information in a human-powered vehicle different in at least one of the human-powered vehicle and a rider of the human-powered vehicle.