🌱 Genetic Programming

Genetic programming can automatically discover mathematical expressions (programs) that describe a dataset, a task known as symbolic regression. Unlike other regression methods, GP produces interpretable expressions that can be analyzed and validated by domain experts.

Research on GP for function learning has examined how the choice of function set, terminal set, and evolutionary operators influences the ability of GP to discover compact and accurate expressions from noisy data. Topics include bloat control, parsimony pressure, and the relationship between expression complexity and generalization performance.

Investigation of different representations for genetic programming, including tree-based, linear, and graph-based representations. Research on variation operators (crossover and mutation) that preserve program semantics, reduce bloat, and improve search efficiency. Development of semantics-based operators that respect the functional behavior of programs rather than operating purely on their syntactic structure.

Grammatical evolution (GE) is a variant of GP that uses a formal grammar (typically in BNF format) to specify the language of programs to be evolved. GE maps a binary or integer string genotype to a program phenotype through grammar-guided derivation. This approach allows the search space to be constrained to syntactically valid programs and enables the evolution of programs in arbitrary programming languages.

Application of grammar-based genetic programming to the automatic design of deep neural network architectures. A formal grammar defines the space of valid network architectures, and genetic programming explores this space to find architectures optimized for specific tasks such as image segmentation and classification.

Development of grammar-based GP methods for automatically designing convolutional neural network architectures. The grammar encodes design choices such as the types of layers, the number of filters, kernel sizes, and connectivity patterns. The evolutionary search finds architectures that achieve high accuracy while remaining computationally efficient.

Research on using GP within evolutionary methods for designing multi-network architectures for heterogeneous multi-task learning. GP provides a flexible representation for specifying how tasks share neural network components, allowing the evolutionary search to discover efficient parameter sharing schemes.

Contribution to the Handbook of Evolutionary Machine Learning: a chapter dedicated to evolutionary methods for unsupervised learning tasks, covering evolutionary approaches to clustering, dimensionality reduction, and generative modeling. The chapter surveys the state of the art and identifies open research questions in evolutionary unsupervised learning.

Contribution to the IEEE CIS Open Book "Introduction to Computational Intelligence": a chapter covering other search-based optimization approaches, including genetic programming, simulated annealing, ant colony optimization, and particle swarm optimization, with an emphasis on their connections to evolutionary computation.

• Santana R (2017). Reproducing and learning new algebraic operations on word embeddings using genetic programming. GECCO 2017.
• Santana R (2021). Semantic Composition of Word-Embeddings with Genetic Programming. GECCO 2021.
• Lima R, Santana R and Pozo A (2019). Automatic design of convolutional neural networks using grammatical evolution. BRACIS 2019.
• Lima R, Santana R and Pozo A (2021). Automatic Design of Deep Neural Networks Applied to Image Segmentation Problems. Applied Soft Computing.
• Lima R, Santana R and Pozo A (2022). A grammar-based GP approach applied to the design of deep neural networks. GECCO 2022.
• Santana R, Mendiburu A and Lozano JA (2019). GP-based methods for domain adaptation: Using brain decoding across subjects as a test-case. GECCO 2019.
• Fontoura VD, Pozo AT and Santana R (2017). Automated design of hyper-heuristics components to solve the PSP problem with HP model. CEC 2017.
• Roman I, Santana R, Mendiburu A and Lozano JA (2019). Sentiment analysis with genetically evolved Gaussian kernels. GECCO 2019.
• Roman I, Santana R, Mendiburu A and Lozano JA (2021). Evolution of Gaussian Process kernels for machine translation post-editing effort estimation. Applied Soft Computing.