Server-Based Device Failure Reproduction via Minimal Action Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies are inefficient in determining the root cause of device issues, leading to impractical and often fruitless efforts in reproducing and solving catastrophic bugs, with many severe problems escaping quality assurance testing and only being discovered when they affect customers in the field.
Innovation Solution
A server creates a reproduction environment to iteratively refine a minimal model using machine learning techniques, combining reinforcement learning and unsupervised learning to determine the minimal set of actions required to reproduce a device condition, allowing for the automatic recreation of device issues and their root causes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual methods are used to reproduce device conditions, then engineers can attempt to understand the problem, but the process becomes extremely time-consuming and often unsuccessful
Solution Approach 1:
The system creates a digital model (solution model) that copies and represents the complex sequence of actions leading to the device condition. Instead of manually reproducing thousands of steps, the system generates a simplified model that captures the essential reproduction path, enabling efficient and repeatable reproduction of the device condition without requiring extensive manual effort.
Solution Approach 2:
The patent replaces the mechanical/manual process of reproducing device conditions with an automated machine learning system. The ML model automatically generates and executes sequences of actions to reproduce the target device condition, substituting human engineers' manual efforts with an automated computational system that can explore action spaces systematically and efficiently.
2Reliability
If comprehensive testing is performed to catch severe problems, then quality assurance improves, but the complexity and cost of testing increases significantly
Solution Approach 1:
The system changes the parameters of the testing approach by using machine learning to intelligently select and prioritize test actions based on the solution model. Instead of exhaustive testing of all possible action sequences, the ML model focuses on the specific parameter space relevant to reproducing the target condition, reducing testing complexity while maintaining reliability.
Solution Approach 2:
The solution model serves as a copied representation of the reproduction process, allowing the system to simulate and analyze device conditions without requiring complex physical test setups. The digital model enables comprehensive testing scenarios to be executed virtually, reducing the complexity of the physical testing infrastructure needed.
3Loss of information
If engineers manually analyze device failures to determine root cause, then they can identify problems, but the process is inefficient and often fruitless for rare conditions
Solution Approach 1:
The patent replaces manual analysis with an automated machine learning system that systematically explores action sequences to determine root causes. The ML model can process and analyze far more data points and action combinations than human engineers, dramatically improving productivity in root cause identification while maintaining or enhancing the quality of analysis through systematic exploration of the action space.
Solution Approach 2:
The system enables self-service problem solving by automatically generating solution models that identify root causes without requiring extensive human intervention. The ML model autonomously analyzes device failures, generates reproduction sequences, and identifies root causes, allowing the system to serve itself in the problem-solving process and freeing engineers from manual analysis tasks.
Data Source
AI summary
In one embodiment, a server creates a reproduction environment of a particular condition of a particular device, the reproduction environment having a device under test (DUT) representative of the particular device, and also being seeded with features regarding the particular condition. The server generates a plurality of models for reaching a target state of the particular condition, each of the plurality of models having differing actions. According to the techniques herein, the server then iteratively refines a minimal model based on the actions of the plurality of models and whether those actions during testing of the DUT get closer to or further from the target state. In response to determining that the minimal model can no longer be further refined during the iterative refining, the server then stores the minimal model as a solution model indicating a given minimal set and order of actions required to reach the target state.