Neural Network Reinforcement Testing via Neuro-Evolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network reinforcement testing methods require human interaction and are limited in their ability to autonomously learn and adapt to new test cases and regression buckets over multiple iterations.
Innovation Solution
A computer-implemented method for reinforcement testing of neural networks using neuro-evolution of augmenting topologies (NEAT) that iteratively creates, evaluates, and adjusts neural network instantiations based on fitness scores, allowing the network to automatically learn and improve without human intervention by executing tests, evaluating states, and selecting mutations to refine performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional neural network testing methods are used, then human interaction is required to evaluate test cases, but this reduces productivity and increases time consumption
Solution Approach 1:
The neural network is configured to autonomously evaluate test cases and determine fitness scores without human intervention. The system self-services by automatically executing test cases, analyzing results, and iteratively improving its own performance through the fitness function, eliminating the need for human evaluators and significantly increasing productivity.
Solution Approach 2:
The system implements a feedback mechanism where test results are automatically fed back into the neural network to adjust weights and improve future predictions. This closed-loop feedback system enables continuous self-improvement and eliminates the need for repeated human evaluation, reducing time loss while maintaining high productivity.
2Adaptability or versatility
If manual testing methods are used, then test cases can be evaluated, but the system cannot autonomously learn and adapt to new test cases and regression buckets
Solution Approach 1:
The neural network autonomously learns from test results by automatically adjusting its weights based on fitness scores. The system services itself by independently identifying patterns in new test cases and regression buckets, continuously improving its adaptability without requiring external retraining or human intervention, thereby achieving both high adaptability and automation.
Solution Approach 2:
The patent replaces manual mechanical testing processes with an automated computational system. The neural network uses algorithmic mechanisms to automatically evaluate, learn from, and adapt to new test cases, substituting human-operated mechanical testing with an autonomous digital system that achieves superior adaptability and automation.
3Measurement precision
If fitness scores are calculated for all instantiations, then optimal solutions can be identified, but this increases computational complexity and processing time
Solution Approach 1:
The system dynamically adjusts parameters such as population size, mutation rates, and fitness function thresholds based on iteration progress. By changing these parameters adaptively, the system maintains high measurement precision in fitness score calculation while reducing computational complexity in later iterations when convergence is approached, balancing accuracy with efficiency.
Solution Approach 2:
The system calculates fitness scores for a selected subset of instantiations rather than all possible instantiations at each iteration. This partial action approach maintains sufficient measurement precision to identify optimal solutions while significantly reducing computational complexity and processing time by focusing calculations on the most promising candidates.
Data Source
AI summary
Aspects of the invention include creating a neural network including neurons to which actions are assigned in representation of test cases. Tests of various instantiations of the neural network are executed for each test case and a state of the neural network after each test is evaluated to determine a fitness score of a corresponding instantiation. Instantiations having fitness scores that exceed a predefined level and identified and the instantiations having the fitness scores that exceed the predefined level are selected for adjustments. The executing, the evaluating, the identifying and the selecting are iteratively repeated in order to obtain desired fitness scores.