- author = "Luigi Rovito",
- title = "Genetic Programming Across Domains: Leveraging Evolutionary Computation to Address Practical Problems",
- school = "Department of Mathematics, Informatics, and Geosciences, University of Trieste",
- year = "2024",
- address = "Italy",
- month = oct,
- keywords = "genetic algorithms, genetic programming, Genetic Improvement, Grammatical Evolution, AI, XAI, ANN",
-
URL = "
https://arts.units.it/handle/11368/3106312",
-
URL = "
https://arts.units.it/retrieve/59caa1e1-7000-468f-a8f3-7556a2dfc67e/PhDThesisLuigiRovito.pdf",
- size = "164 pages",
-
abstract = "Several modern technologies are based on Artificial
Intelligence (AI) techniques, which are continuously
studied and refined in many research fields. Among the
subfields of AI, bio-inspired approaches, such as
Genetic Programming (GP), have found great interest
because of their effectiveness in discrete optimization
problems that are prevalent in fields such as
cryptography, code improvement, and Interpretable
Machine Learning (IML). In this work, we investigate
the application of GP methods in the three
aforementioned topics.
Firstly, we employ GP to discover non-linear Boolean functions by exploring the search space of their Walsh transform-based representations for stream ciphers design. Boolean functions are essential components of secure cryptographic systems, and we specifically focus on optimizing non-linearity, a key security property, as well as balancedness, which provides resistance to statistical attacks. The proposed method demonstrates the capability of GP to evolve highly secure Boolean functions with a large number of variables, which are typically used in practical cryptographic scenarios.
Secondly, we try to improve the correctness of code generated by a Large Language Model (LLM) by leveraging a Genetic Improvement (GI) approach that, internally, employs a GP method defined as Grammatical Evolution (GE). This method starts from the code generated by an LLM, evaluates its correctness based on a set of user-provided test cases, and iteratively evolves improved solutions. Our results highlight the potential of combining GI and GE to enhance the reliability and correctness of automatically generated code, thus addressing one of the key limitations of current automatic code generation tools.
Thirdly, we propose a human-in-the-loop GP framework to evolve generic tree-based Machine Learning (ML) models that are evaluated in terms of both qualitative performance and interpretability. The interpretability of models is estimated by an Artificial Neural Network (ANN) that is trained with user feedback to capture the subjectivity of the user. This framework not only enables the discovery of models with competitive predictive performance but also provides users with interpretable solutions tailored to their specific background and requirements, which is a critical need in high-stakes applications.
Finally, we study how GP itself can be improved on its most famous problem, that is, Symbolic Regression (SR). Specifically, we extend a GP variant called Geometric Semantic Genetic Programming (GSGP) to develop Cellular Geometric Semantic Genetic Programming (cGSGP), which improves the diversity of the solutions evolved during the optimization process by imposing a neighbourhood-based toroidal structure over the population. By limiting genetic operations to individuals in the same local neighborhood, cGSGP mitigates premature convergence and improves the exploration of the search space, leading to the discovery of more accurate symbolic models.
Our results highlight the effectiveness of GP in tackling the discussed problems, especially as regards building interpretable ML models, which is an urgent problem in high-stakes applications where the transparency of decisions is critical",
-
notes = "in English ING-INF/05
Supervisor: Andrea De Lorenzo",
