A fundamental problem in science is the formulation of a model of a system when no a priori knowledge of the process is available. To this end, symbolic regression methods were developed to search for a model’s structure, as well as its parameters, using a stochastic optimization technique known as genetic programming. However, symbolic regression exhibits a well-known phenomenon known as bloat, in which it produces overly complex solutions. We developed a method known as epigenetic local search [1] that specifically optimizes model structures for conciseness during optimization via a novel developmental encoding. We found that this method produced concise solutions that generalized better to test data. We applied it to the modeling of wind turbine dynamics [2], as well as fluid dynamics and other non-linear dynamic systems [3] and showed it could find true underlying physical principles. In comparison to neural network and auto-regressive models, our method produced much simpler and often more accurate models.

  1. La Cava, W., Helmuth, T., Spector, L., & Danai, K. (2015). Genetic Programming with Epigenetic Local Search. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO) (pp. 1055–1062). ACM Press. link, preprint

  2. La Cava, W., Danai, K., & Spector, L. (2016). Inference of compact nonlinear dynamic models by epigenetic local search. Engineering Applications of Artificial Intelligence, 55, 292–306. link, preprint

  3. La Cava, W., Danai, K., Spector, L., Fleming, P., Wright, A., & Lackner, M. (2016). Automatic identification of wind turbine models using evolutionary multiobjective optimization. Renewable Energy, 87, Part 2, 892–902. link, preprint