Forecasting Approximation Architecture with Dimensionality Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural Networks (NNs) face challenges in the learning phase due to high computational effort and cost, as well as accuracy issues, particularly in forecasting systems where repetitive evaluation of computationally expensive functions is slow and costly, making practical applications impractical.
Innovation Solution
A method for training an approximation architecture in forecasting systems that reduces computational load by identifying dependencies between input variables, applying dimensionality-reducing techniques to generate modified training samples, and optimizing the architecture to require fewer resources for accurate approximation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional neural network training is used for forecasting systems, then the model can learn complex functions, but the computational effort and cost during the learning phase becomes prohibitively high
Solution Approach 1:
The patent applies preliminary action by performing dimensionality reduction on the training data before neural network training. The method identifies and exploits dependencies between input variables to transform the original high-dimensional training samples into lower-dimensional representations. This pre-processing step reduces the computational complexity of subsequent training operations while preserving the essential information needed for accurate function approximation in forecasting systems
Solution Approach 2:
The patent extracts the essential dependencies between input variables from the training data and removes redundant dimensions. By identifying which variables are truly independent and which can be derived from others, the method extracts only the necessary degrees of freedom needed for training, thereby reducing computational cost while maintaining approximation accuracy
2Measurement precision
If traditional neural network training is used for forecasting systems, then the model can learn complex functions, but the training time becomes excessively long
Solution Approach 1:
The patent performs dimensionality reduction as a preliminary action before training begins. By transforming training samples to lower dimensions based on variable dependencies, the method reduces the number of iterations and computational operations required during training, thereby significantly reducing training time while preserving the ability to learn complex forecasting functions
Solution Approach 2:
The patent changes the parameter space by reducing the dimensionality of input variables. The method transforms the original set of input variables into a reduced set that captures the essential relationships, thereby changing the parameters of the training problem to reduce computational complexity and training time
3Device complexity
If homogeneous neural network architectures are used, then the system is simpler to implement, but the computational cost of training remains prohibitively high for forecasting applications
Solution Approach 1:
The patent applies preliminary dimensionality reduction to simplify the training problem before training begins. By reducing the input dimensionality based on variable dependencies, the method makes training computationally feasible for homogeneous architectures without requiring complex architectural modifications, thereby reducing training computational cost while maintaining implementation simplicity
Data Source
AI summary
The invention provides a computer-implemented method of training an approximation architecture for a forecasting system including performing a pre-training stage comprising steps of: obtaining training samples each including an input value for a first plurality of input variables and corresponding parameter values of the functions; obtaining dependencies between different variables of the first plurality of input variables; determining, based on the obtained dependencies, dimensionality-reducing rules for determining a second plurality of input variables, wherein there are fewer degrees of freedom in the second plurality than in the first plurality; and, determining, by applying the dimensionality-reducing rules to the training samples, modified training samples including input values for at least some of the second plurality of variables and corresponding parameter values approximating the functions. The forecasting system is for determining a future behaviour forecast of a system based on determined approximated parameter values, from the trained approximation architecture, for the plurality of scenarios.