Description

This course explores the intersection of evolutionary algorithms and neural networks, focusing on modern approaches to evolving neural architectures, weights, and learning algorithms. Students will study how evolutionary computation can complement or replace gradient-based methods, particularly in scenarios where gradients are unavailable or ineffective. The course covers evolution strategies, genetic programming, neuroevolution techniques, and neural architecture search, emphasizing both theoretical foundations and practical applications. Students will understand how evolutionary approaches enable automated machine learning, open-ended learning, and the discovery of novel neural architectures.

Students could implement evolution strategies for optimization problems, conduct neural architecture search projects comparing evolutionary and gradient-based approaches, use genetic programming for automated feature engineering, develop hybrid systems combining evolution with gradient descent, and complete a final project applying evolutionary AI to a domain such as robotics control, game playing, or automated model design.

Core Topics

The course begins with advanced evolutionary computation theory beyond M1S2, including evolution strategies (CMA-ES, Natural ES), genetic programming with tree-based representations, multi-objective evolutionary algorithms, and quality diversity approaches like novelty search. Students then study neuroevolution fundamentals, including methods for evolving neural network weights through direct encoding, the NEAT algorithm for evolving topologies, HyperNEAT for indirect encodings, and compositional pattern-producing networks. A major focus is neural architecture search, covering search space design, evolutionary NAS versus gradient-based approaches, efficient NAS methods with weight sharing, hardware-aware NAS, and multi-objective optimization balancing accuracy with efficiency. The course explores hybrid approaches that combine evolution with gradient descent, evolutionary strategies for reinforcement learning, and Lamarckian evolution in neural networks. Advanced topics include automated machine learning (AutoML), evolving activation functions and optimizers, genetic programming for symbolic regression, and open-ended evolution. Finally, students learn about parallel and distributed evolution, including island models, massively parallel fitness evaluation, GPU-accelerated evolutionary algorithms, and distributed NAS systems.

MCC

Les épreuves indiquées respectent et appliquent le règlement de votre formation, disponible dans l'onglet Documents de la description de la formation.

Régime d'évaluation

ECI (Évaluation continue intégrale)

Coefficient

3.0

Évaluation initiale / Session principale - Épreuves