Genetic Programming and Evolvable Machines最新文献

Geometric semantic genetic programming (GSGP) represents one of the most promising developments in the area of evolutionary computation (EC) in the last decade. The results achieved by incorporating semantic awareness in the evolutionary process demonstrate the impact that geometric semantic operators have brought to the field of EC. An improvement to the geometric semantic mutation (GSM) operator is proposed, inspired by the results achieved by batch normalization in deep learning. While, in one of its most used versions, GSM relies on the use of the sigmoid function to constrain the semantics of two random programs responsible for perturbing the parent’s semantics, here a different approach is followed, which allows reducing the size of the resulting programs and overcoming the issues associated with the use of the sigmoid function, as commonly done in deep learning. The idea is to consider a single random program and use it to perturb the parent’s semantics only after standardization or normalization. The experimental results demonstrate the suitability of the proposed approach: despite its simplicity, the presented GSM variants outperform standard GSGP on the studied benchmarks, with a difference in terms of performance that is statistically significant. Furthermore, the individuals generated by the new GSM variants are easier to simplify, allowing us to create accurate but significantly smaller solutions.

Linear genetic programming (LGP) is a genetic programming paradigm based on a linear sequence of instructions being executed. An LGP individual can be decoded into a directed acyclic graph. The graph intuitively reflects the primitives and their connection. However, existing studies on LGP miss an important aspect when seeing LGP individuals as graphs, that is, the reverse transformation from graph to LGP genotype. Such reverse transformation is an essential step if one wants to use other graph-based techniques and applications with LGP. Transforming graphs into LGP genotypes is nontrivial since graph information normally does not convey register information, a crucial element in LGP individuals. Here we investigate the effectiveness of four possible transformation methods based on different graph information including frequency of graph primitives, adjacency matrices, adjacency lists, and LGP instructions for sub-graphs. For each transformation method, we design a corresponding graph-based genetic operator to explicitly transform LGP parent’s instructions to graph information, then to the instructions of offspring resulting from breeding on graphs. We hypothesize that the effectiveness of the graph-based operators in evolution reflects the effectiveness of different graph-to-LGP genotype transformations. We conduct the investigation by a case study that applies LGP to design heuristics for dynamic scheduling problems. The results show that highlighting graph information improves LGP average performance for solving dynamic scheduling problems. This shows that reversely transforming graphs into LGP instructions based on adjacency lists is an effective way to maintain both primitive frequency and topological structures of graphs.

Progress in AI has brought new approaches for designing products via co-creative human–computer interaction. In architecture, interior design, and industrial design, computational methods such as evolutionary algorithms support the designer’s creative process by revealing populations of computer-generated design solutions in a parametric design space. Because the benefits and shortcomings of such algorithms’ use in design processes are not yet fully understood, the authors studied the intricate interactions of an industrial designer employing an interactive evolutionary algorithm for a non-trivial creative product design task. In an in-depth report on the in-situ longitudinal experiences arising between the algorithm, human designer, and environment, from ideation to fabrication, they reflect on the algorithm’s role in inspiring design, its relationship to fixation, and the stages of the creative process in which it yielded perceived value. The paper concludes with proposals for future research into co-creative AI in design exploration and creative practice.

Boolean functions are important cryptographic primitives with extensive use in symmetric cryptography. These functions need to possess various properties, such as nonlinearity to be useful. The main limiting factor of the quality of a Boolean function is the number of its input variables, which has to be sufficiently large. The contemporary design methods either scale poorly or are able to create only a small subset of all functions with the desired properties. This necessitates the development of new and more efficient ways of Boolean function design. In this paper, we propose a new semantic mutation operator for the design of bent Boolean functions via genetic programming. The principle of the proposed operator lies in evaluating the function’s nonlinearity in detail to purposely avoid mutations that could be disruptive and taking advantage of the fact that the nonlinearity of a Boolean function is invariant under all affine transformations. To assess the efficiency of this operator, we experiment with three distinct variants of genetic programming and compare its performance to three other commonly used non-semantic mutation operators. The detailed experimental evaluation proved that the proposed semantic mutation operator is not only significantly more efficient in terms of evaluations required by genetic programming but also nearly three times faster than the second-best operator when designing bent functions with 12 inputs and almost six times faster for functions with 20 inputs.

Evolutionary feature construction has been successfully applied to various scenarios. In particular, multi-tree genetic programming-based feature construction methods have demonstrated promising results. However, existing crossover operators in multi-tree genetic programming mainly focus on exchanging genetic materials between two trees, neglecting the interaction between multi-trees within an individual. To increase search effectiveness, we take inspiration from the geometric semantic crossover operator used in single-tree genetic programming and propose a macro geometric semantic crossover operator for multi-tree genetic programming. This operator is designed for feature construction, with the goal of generating offspring containing informative and complementary features. Our experiments on 98 regression datasets show that the proposed geometric semantic macro-crossover operator significantly improves the predictive performance of the constructed features. Moreover, experiments conducted on a state-of-the-art regression benchmark demonstrate that multi-tree genetic programming with the geometric semantic macro-crossover operator can significantly outperform all 22 machine learning algorithms on the benchmark.