Machine learning device, machine learning method, and storage medium
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The efficiency of relearning in air conditioning control systems is low due to the generation of machine learning models that are not suitable for individual edges, leading to user discomfort as the models are learned using mixed air conditioning control environments and may not adapt accurately, resulting in repeated relearning cycles.
Innovation Solution
A method where a cloud server distributes a machine learning model learned from multiple edges, identifies and excludes data from edges that are not suitable for a specific edge, and relearns the model using the adjusted training data, ensuring the model is adapted to each edge's unique environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a machine learning model is learned using mixed air conditioning control environments from multiple edges, then the model can be distributed to multiple edges, but the model does not adapt accurately to individual edges leading to user discomfort
Solution Approach 1:
The patent segments the training data by edge, creating separate training datasets for each edge device. The learning unit learns multiple machine learning models corresponding to multiple edges respectively, rather than creating a single aggregated model. This segmentation allows each model to be specialized for its target edge while maintaining overall system versatility.
Solution Approach 2:
The patent applies local quality by making each machine learning model have different characteristics suited to its specific edge environment. The learning unit generates models with local adaptations based on edge-specific training data, ensuring each model has the appropriate quality and parameters for its target edge device rather than using a uniform model for all edges.
2Reliability
If relearning is performed frequently to improve model accuracy for each edge, then user comfort improves, but the time and computational resources required increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-learning multiple machine learning models for multiple edges in advance, before actual air conditioning control operations begin. The learning unit learns models corresponding to each edge respectively using collected training data, so that when control operations start, accurate models are already available, eliminating the need for frequent relearning during operation.
3Quantity of substance
If a machine learning model is learned using data from all edges, then more training data is available, but the model becomes less suitable for individual edges requiring repeated relearning
Solution Approach 1:
The patent segments the training data by edge, creating separate training datasets for each edge device. The learning unit learns multiple machine learning models corresponding to multiple edges respectively, rather than creating a single aggregated model. This segmentation allows each model to be specialized for its target edge while maintaining overall system versatility.
Solution Approach 2:
The patent applies parameter changes by adjusting the training data parameters for each edge-specific model. The learning unit learns models with different parameters and characteristics suited to each edge's specific environment, using edge-specific training data rather than a uniform parameter set for all edges.
Data Source
AI summary
A machine learning method executed by a computer, the method includes distributing a first learning model learned on the basis of a plurality of logs collected from a plurality of electronic devices to each of the plurality of electronic devices, the first learning model outputting operation content for operating an electronic device; when an operation different from an output result of the first learning model is performed by a user relative to a first electronic device among the plurality of electronic devices, estimating a similar log corresponding to a state of the learning model in which the different operation is performed from the plurality of logs; generating a second learning model on the basis of a log obtained by excluding a log of a second electronic device associated with the similar log from among the plurality of logs; and distributing the second learning model to the first electronic device.