- author = "Luiz Otavio Vilas Boas Oliveira",
- title = "Improving Search in Geometric Semantic Genetic Programming",
- school = "Universidade Federal de Minas Gerais",
- year = "2016",
- address = "Belo Horizonte, Brazil",
- month = "28 " # sep,
- keywords = "genetic algorithms, genetic programming, Geometric Semantic Genetic Programming, GSGP, GSX, LGX, AGD, MGD, Programacao genetica (Computacao), Computacao, Operadores semanticos geometricos, Operadores de dispersao, selecao de instancias, Programacao genetica sementica",
- timestamp = "2016.09.27",
- owner = "luiz",
- file = "Improving Search in Geometric Semantic Genetic Programming.pdf:Artigos/my-publications/Improving Search in Geometric Semantic Genetic Programming.pdf:PDF",
-
URL = "
https://luizvbo.github.io/publication/oliveira-2016-improving/cite.bib",
-
URL = "https://luizvbo.github.io/publication/oliveira-2016-improving/",
-
URL = "https://repositorio.ufmg.br/items/1e17cc72-4fb3-4e78-a1fc-01c0c17677fc",
-
URL = "https://hdl.handle.net/1843/ESBF-AFWG3P",
-
URL = "
https://repositorio.ufmg.br/bitstream/1843/ESBF-AFWG3P/1/luizotaviovilasboasoliveira.pdf",
- size = "147 pages",
-
abstract = "In the last few years a variety of works has emerged
attempting to introduce semantic awareness to the
evolutionary process performed by Genetic Programming
(GP). In particular, Geometric Semantic GP (GSGP) was
proposed as a method that acts on the syntax of the
parent programs producing offspring respecting a
semantic criterion.
In this thesis we focus on the open issues presented by GSGP and related methods, applied to symbolic regression. We present the definition of semantics adopted in this thesis and an overview of the methods that explore semantics framed in this definition. Then we present and tackle issues regarding GSGP.
We first investigate the impact of the geometric semantic crossover with different distance functions on the search, and the possibility of optimally adjusting its coefficients instead of choosing them randomly. The results show that the Manhattan distance has best performance in terms of test error, and that optimizing the crossover coefficients cannot improve significantly the search.
We also present the Sequential Symbolic Regression (SSR), an attempt to control the exponential growth on the size of the individuals caused by the use of the geometric semantic crossover. After generating a suboptimal function with a canonical GP, SSR approximates the output errors by another function, in a subsequent iteration, and concatenates them with the crossover operator. An experimental analysis shows that SSR has similar performance to GSGP while generating smaller solutions.
In addition, this thesis explores a heuristic framework, called Geometric Dispersion (GD), to construct operators that move individuals to less dense areas of the search space around the target output vector. Experimental results indicate that GD operators can improve the search and spread the solutions around the target solution.
Last, we present a study of the impact of selecting training instances in order to reduce the semantic space dimensionality. Two approaches are considered: (i) to apply current instance selection methods as a pre-process step before training points are given to GSGP; (ii) to incorporate instance selection to the evolution of GSGP. The experimental analysis shows that GSGP performance is improved by using instance reduction during the evolution.",
-
notes = "in english
CDU 519.6*82(043) CNPq support
Supervisor: Gisele Lobo Pappa",
