ED Mathématiques et Informatique
Towards open-ended dynamics in artificial life and artificial intelligence: an eco-evo-devo perspective
by Gautier HAMON (Institut national de recherche en informatique et en automatique - Bordeaux - Sud-Ouest)
The defense will take place at 14h30 - Ada Lovelace 200 Av. de la Vieille Tour, Centre Inria de l'université de Bordeaux, 33405, Talence
in front of the jury composed of
- Clément MOULIN-FRIER - Chargé de recherche - Centre Inria de l'université de Bordeaux - Directeur de these
- Nicolas BREDECHE - Professeur des universités - Sorbonne Université, Paris, France - Rapporteur
- Daniel POLANI - Professeur - University of Hertfordshire - Rapporteur
- Lisa SOROS - Roman Family Teaching and Research Fellow - Barnard College - Examinateur
- Antoine CULLY - Assistant professor - Imperial College London - Examinateur
- Nicolas ROUGIER - Directeur de recherche - Institut des Maladies Neurodégénératives, and Inria - Examinateur
Natural evolution has, over billions of years, gradually generated the astonishing diversity of complex life forms that populates our planet. This phenomenon exemplifies what we call an open-ended process: a system capable of continuously generating increasingly diverse and complex structures. Inspired by this phenomena as well as other open-ended processes such as human developmental learning and cultural evolution, this thesis investigates key mechanisms that underpin open-ended processes and emergent complexity. Situated at the intersection of artificial life, machine learning, and open-endedness, this thesis explores, in simulations, emergent complexity across varying levels of abstraction. We focus on the importance of environment dynamics and its interplay with adapting agents in this quest of open-ended dynamics in silico. In particular, we highlight the major effects of feedback loop dynamics, such as co-adaptation in a group of agents or agent-environment reciprocal causation -- wherein agents adapt to the environment but also alter it through their own behavior, which in turn modify the environment and shape their adaptation. For this aim, we rely on diverse state-of-art methods from artificial life and machine learning, including cellular automata, diversity search, neuroevolution, multi-agent systems and meta reinforcement learning. The thesis explores emergent complexity at different levels of abstraction. First, it explores the genesis of individuality within an originally lifeless simulated environment composed of simple atomic elements and local physical rules, also probing how such environments can bootstrap evolutionary dynamics. Next, assuming the existence of agents and evolutionary processes, the focus shifts to how adapting agents actively modify their environments -- potentially to their advantage --thereby altering evolutionary pressures. These new pressures, in turn, influence the agents' subsequent adaptations and therefore actions on the environment, creating feedback loops that perpetually drive new adaptations in a potentially open-ended way. Finally, the research explores how these continual environmental changes may foster the development of faster adaptation mechanisms, enabling agents to cope with this high environmental variability. Specifically, we examine how variable environments can facilitate the emergence of efficient exploratory behaviors within groups of agents. By investigating these phenomena, this research contributes foundational insights toward designing systems capable of bootstrapping and sustaining open-ended processes, ultimately reflecting the rich, adaptive complexity of the natural world --from the origins of life to the evolution of generalist agents.
ED Sciences Physiques et de l'Ingénieur
Design and FPGA prototyping of an advanced receiver based on expectation propagation
by Ian FISCHER SCHILLING (Laboratoire de l'Intégration du Matériau au Système)
The defense will take place at 10h15 - Amphi J. P. DOM A0.85 351 Cours de la Libération, Laboratoire IMS, Batiment A31, 33405 Talence Cedex, France
in front of the jury composed of
- Christophe JEGO - Professeur des universités - ENSEIRB-MATMECA - Bordeaux INP - Directeur de these
- Jean-Pierre CANCES - Professeur des universités - ENSIL-ENSCI - Université de Limoges - Rapporteur
- Raphaël LE BIDAN - Maître de conférences - IMT Atlantique - Rapporteur
- Charly POULLIAT - Professeur des universités - ENSEEIHT - Toulouse INP - Examinateur
- Antonio CIPRIANO - Docteur - Thales SIX GTS - Examinateur
- Camille LEROUX - Maître de conférences - ENSEIRB-MATMECA - Bordeaux INP - CoDirecteur de these
Expectation Propagation (EP) is a powerful technique used in statistical inference to approximate complex probability distributions with simpler ones from the exponential family through moment matching. Recent works have demonstrated that its application in digital receiver design offers an attractive complexity-performance trade-off. By iteratively refining signal estimates via a message-passing approach, EP provides a robust framework for addressing challenges in digital communication systems, such as inter-symbol interference (ISI) in wideband channels. In this thesis, an EP-based Frequency Domain Self-Iterated Linear Equalizer (FD-SILE) is considered, comprising an equalizer, a soft demapper and a soft mapper. These components take advantage of EP for feedback within a self-iterating process. While the EP-based FD-SILE demonstrates favorable complexity-performance, its computational complexity remains prohibitive for hardware implementations, particularly for high-order constellations. In order to decrease this computational complexity, analytical simplifications are introduced for the soft mapping and demapping processes. These simplifications achieve substantial reductions in computational complexity while preserving bit error rate (BER) performance. As part of this thesis work, fixed-point versions of the simplified soft mapper and demapper are carried out to enable architecture design. Different architectures are designed for the modulation schemes of BPSK, QPSK, 8-PSK, and 16-QAM. These architectures are then optimized through pipelining, significantly reducing the number of clock cycles per frame. A flexible pipelined architecture, capable of dynamically switching constellations on a per-frame basis, is subsequently designed and implemented onto an FPGA device. Validation is conducted using a hardware-in-the-loop (HIL) configuration, which integrates a simulation environment on a computer with the FPGA-implemented architecture on a Zynq MPSoC platform.