System and Method for Discovering Optimised Combinations of Calculation Functions

Simple SummaryContent extracted from patent full text and abstract with AI.

This invention provides a system and method that automates the discovery of optimized combinations of calculation functions to extract the most relevant features from data for machine learning applications. It uses a reinforcement learning agent that experiments with different combinations of feature-calculation methods, evaluates their utility in improving a machine learning task (like classification), and iteratively converges to the most effective and efficient set of feature calculations. The system can automatically switch between training and runtime modes and adapt to changes in the data or environment.

Use CasesContent extracted from patent full text and abstract with AI.

• Automated feature engineering in data science workflows across industries (e.g., finance, healthcare, marketing).
• Optimizing preprocessing pipelines for machine learning models in industrial monitoring and quality assurance.
• Automated analysis and classification of sensor data in IoT applications.
• Medical data analysis, where optimal feature combinations can aid in diagnostics or patient monitoring.
• Smart city applications, such as traffic analysis and weather prediction.
• Image or signal processing, for automated selection and combination of feature extraction techniques.

BenefitsContent extracted from patent full text and abstract with AI.

• Substantially reduces the manual effort and expert knowledge needed for optimal feature selection and engineering.
• Enables data-driven, adaptive, and context-specific feature extraction, improving model accuracy and performance.
• Can optimize not only for accuracy but also for computational efficiency, important for edge devices or resource-constrained environments.
• Adapts dynamically to changing data distributions or problem domains via its inherent reinforcement learning loop.
• Minimizes trial-and-error and speeds up the development and deployment of high-performing machine learning solutions.
• Reduces risk of sub-optimal model performance due to manual or heuristic feature engineering.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

The invention relates to a system (100) for discovering one or more optimised combination(s) of calculation functions (f1, f2,..., fN) for calculating features (m1, m2,..., mN) of an entity, comprising a learning reinforcement agent (200) which is designed to choose one or more combinations of calculation functions (f1, f2,..., fN) which are forwarded to a data preprocessing module (300), the data preprocessing module (300) being designed to calculate the features (m1, m2,..., mN) from the data of a training data set (320) by means of the selected calculation functions (f1, f2,..., fN) and to forward the features (m1, m2,..., mN) to a machine learning module (400) which is designed to analyse and/or to train the features (m1, m2,..., mN) and to forward the learning result to an evaluation module (440) which is designed to create an evaluation for the learning result and to forward said evaluation to the learning reinforcement agent (200) which is designed to again choose one or more combination(s) of calculation functions (f1, f2,..., fN) on the basis of said evaluation of the learning result.