Accurate High Performance Concrete Prediction with an Alignment-Based Genetic Programming System

Kristen Scott,Leonardo Vanneschi,Mauro Castelli,Aleš Popovič

doi:10.1186/s40069-018-0300-5

Abstract

In 2013, our research group published a contribution in which a new version of genetic programming, called Geometric Semantic Genetic Programming (GSGP), was fostered as an appropriate computational intelligence method for predicting the strength of high-performance concrete. That successful work, in which GSGP was shown to outperform the existing systems, allowed us to promote GSGP as the new state-of-the-art technology for high-performance concrete strength prediction. In this paper, we propose, for the first time, a novel genetic programming system called Nested Align Genetic Programming (NAGP). NAGP exploits semantic awareness in a completely different way compared to GSGP. The reported experimental results show that NAGP is able to significantly outperform GSGP for high-performance concrete strength prediction. More specifically, not only NAGP is able to obtain more accurate predictions than GSGP, but NAGP is also able to generate predictive models with a much smaller size, and thus easier to understand and interpret, than the ones generated by GSGP. Thanks to this ability of NAGP, we are able here to show the model evolved by NAGP, which was impossible for GSGP.

Highlights

Concrete is a material mainly used for the construction of buildings
Fitness is often calculated by running the program on a set of data and quantifying the difference between the behaviour of the program on those data and the expected behaviour. It is a recent trend in the Genetic Programming (GP) research community to develop and study methods to integrate semantic awareness in this evolutionary process, where with the term semantics we generally indicate the vector of the output values calculated by a model on all the available training cases
Our work from 2013 (Castelli et al 2013) clearly indicated that a relatively recent and sophisticated version of GP, that exploits semantic awareness, called Geometric Semantic GP (GSGP) (Moraglio et al 2012; Vanneschi 2017) is able to outperform standard GP, predicting high-performance concrete (HPC) strength with high accuracy. This success motivated us to pursue the research, with the objective of further improving the results obtained by Geometric Semantic Genetic Programming (GSGP), possibly investigating new and more promising ways of exploiting semantic awareness

Summary

Background

Concrete is a material mainly used for the construction of buildings. It is made from a mixture of broken stone or gravel, sand, cement, water and other possible aggregates, which can be spread or poured into moulds and forms a stone-like mass on hardening. Our work from 2013 (Castelli et al 2013) clearly indicated that a relatively recent and sophisticated version of GP, that exploits semantic awareness, called Geometric Semantic GP (GSGP) (Moraglio et al 2012; Vanneschi 2017) is able to outperform standard GP, predicting HPC strength with high accuracy This success motivated us to pursue the research, with the objective of further improving the results obtained by GSGP, possibly investigating new and more promising ways of exploiting semantic awareness. We define a novel computational intelligence method, called Nested Align Genetic Programming (NAGP), based on the concept of alignment in the error space, that is able to outperform GSGP for HPC strength prediction The former contribution is important for the GP community because it represents a further step forward in a very popular research line (improving GP with semantic awareness).

Issue 1

Issue 3

Experimental Study

Findings

Conclusions and Future Work