Decomposition Based Hybrid Symbolic Regression with Multiobjective Genetic Programming

Symbolic regression can determine symbolic models to best depict the laws concealing in the historical data. However, the disadvantages over model bloating, blind search and losing diversity often make its solution algorithms time-consuming and unstable. We propose a decomposition strategy to divide the symbolic regression problem into a sequence of simple subproblems. Each subproblem contains a global and a local regression that can be concurrently solved by various regression methods. Representing symbolic regression problems as multiojective optimization models, an algorithm framework based on genetic programming and decomposition is developed, in which the subtasks are solved concurrently and traditional regression techniques are employed aiming at improving the time efficiency, search efficiency and model diversity. To evaluate the feasibility and performance of the proposed framework, experiments and comparisons are conducted.