Ensemble Model Training via Competitive Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for building ensemble models are inefficient and computationally intensive, as they require re-running multiple combinations of hyper-parameters for machine learning/deep learning models, leading to inaccurate predictions, wastage of computational resources, and lack of feature learning across models.
Innovation Solution
The method creates clusters of predictive models, initializes them with random parameters, categorizes them based on accuracy, and uses competitive reinforcement learning to determine optimal parameters, thereby improving model performance and reducing computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple combinations of hyper-parameters are processed by re-running the ML/DL model from the beginning in each epoch, then the model can find optimized parameters, but the process becomes inefficient, time-consuming, and computationally intensive
Solution Approach 1:
The patent applies preliminary action by pre-processing the dataset into multiple batches before training begins. This allows the model to process data incrementally rather than re-running entire epochs multiple times, significantly reducing computational overhead while still enabling comprehensive hyperparameter search through efficient batch processing
Solution Approach 2:
The patent segments the training process by dividing the dataset into multiple batches and processing them in parallel or sequential manner. This segmentation allows independent processing of different data portions, reducing the time required for each epoch and enabling faster hyperparameter optimization without sacrificing accuracy
2Reliability
If all models in an ensemble are developed independently and then combined optimally during deployment, then each model can learn specific features well, but this results in wastage of computational resources, redundancies, and conflict in decision-making
Solution Approach 1:
The patent merges multiple model training processes into a single unified training loop that processes batches across all models simultaneously. This combining approach eliminates redundant computations, reduces computational resource wastage, and prevents decision conflicts by ensuring all models are trained on the same data distributions and hyperparameter settings
Solution Approach 2:
The patent creates a universal training framework that serves multiple models simultaneously with a single training process. This multi-functional approach allows the same training infrastructure to optimize all ensemble models together, reducing overall computational overhead while maintaining the specialized feature learning capabilities of each individual model
Data Source
AI summary
This disclosure relates to method and system for building ensemble models using competitive reinforcement learning (CRL). The method may include creating a plurality of clusters, each including a set of predictive models forming the ensemble model. For each of the plurality of clusters, the method may further include initializing each of the set of predictive models with random values for a set of first parameters to obtain a first accuracy score, categorizing each of the set of predictive models into a first associated category based on a first accuracy score, calculating a first reward for each of the set of predictive models based on the first associated category, and determining a set of second parameters for each of the set of predictive models using CRL to obtain a second accuracy score.