| KEYWORD |
Evolutionary Approach for Core Knowledge in Artificial Intelligence
keywords GENETIC ALGORITHMS, COMBINATORIAL OPTIMIZATION, MACHINE LEARNING, NEURAL NETWORKS
Reference persons STEFANO QUER, GIOVANNI SQUILLERO
External reference persons Alberto Tonda, INRAE, Université Paris-Saclay, Palaiseau, France
Research Groups DAUIN - GR-13 - METODI FORMALI - FM
Thesis type RESEARCH / EXPERIMENTAL, RESEARCH AND DEVELOPMENT
Description Descrizione Modern machine-learning techniques can effectively create predictive models from large amounts of training data. However, state-of-the-art models are often brittle and not interpretable, making them impractical for applications where the consequences of errors can be expensive or fatal. Several research lines tackle this issue, aiming to provide clear, concise explanations for predictions.
This thesis will focus on an evolutionary methodology inspired by core knowledge, a theory that describes human cognition as a small set of innate abilities combined through compositionality. The work will improve an approach recently developed to generate predictive descriptions of the interactions between elements in simple 2D videos. The methodology exploits well-known strategies, such as image segmentation, object detection, and simple laws of physics (kinematics and dynamics), to evolve rules describing classes of objects and their interactions. The approach, developed in Python, has been tested on two classic video games, Pong and Arkanoid. It has identified objects, classes, and rules on those benchmarks, creating a compact, high-level predictive description of the interactions between the puzzle elements.
The approach is the first step in a research line to build AI systems inspired by core knowledge and exploiting evolutionary computation to aggregate the basic information provided by hard-coded algorithms. The thesis will focus on one or more of the following steps:
• Leverage the descriptions created by the approach to evolve optimal video-game plays, offering a potentially more robust alternative to DL-based solutions.
• Experiment with combinations of EAs with other search algorithms, such as Novelty Search.
• Find more free-form structures to aggregate the outputs of the algorithms used as part of the core knowledge provided to the approach.
• Improve the algorithm's speed using a compiled programming language like C++ or Rust.
Required skills Advanced programming skills and basis on data analytics and machine learning.
Notes The thesis is expected to be carried out in Turin or in Paris in collaboration with Prof. Alberto Tonda from the University Paris-Saclay, Palaiseau, France.
Deadline 01/09/2024
PROPONI LA TUA CANDIDATURA