Author: Kulunchakov Andrei GitHub: KuluAndrej

MVR is a project for construction and approximation of nonlinear functions to a given data. Nonlinear functions are represented as superpositions of primitive functions set up by experts. The generation is done via Genetic Programming (composition of mutations, crossovers and random generations). This process is iterative and each iteration consists of two steps:

• Production of new models. The stored models are participate in mutations and crossovers producing new elements of the corresponding functional space.
• Best models selection. This step embodies natural selection. Only the best representatives of current population are passed to the next iterations.

The project currently has two purposes.

• Data fitting.

We have a data and need to reveal underlying dependencies between independent variables and dependent ones.

• Time series classification.

We have a bunch of time series and need to extract its structural representation. Namely, we approximate nonlinear functions to these time series, represent these functions as labeled trees and extract structural features from these trees. Given features are attached to corresponding time series. Resulted representation could be used in tasks of classification, clustering and anomaly detection.