Background: Recent advances in artificial intelligence have led to the development of increasingly sophisticated multi-agent systems that constantly adapt to their changing environment. These systems have wide-ranging applications, from managing traffic networks and power grids to enabling the creation of virtual economies on blockchain networks. However, as they evolve over time, these systems encounter uncertain or adversarial environments that compromise their safety and stability in the long-run. These encounters may lead to dramatic changes in the future evolution or equilibrium states of a given system, often only due to small perturbations in its critical parameters that are hard to control or predict.
Such critical phase transitions are frequently observed both in the laboratory, e.g., during the training of neural networks or simulation of dynamical systems, and in the real-world, for instance, in financial markets, genetics, disease outbreaks, opinion swifts in social networks, or mass adoption of novel technologies. However, despite their prevalence, the dynamics behind such transitions are still not well understood. In fact, their complexity still constitutes a critical bottleneck in the ability of scientists and policymakers to increase the adoption of these novel technologies and to make the underlying AI systems more safe, trustworthy, fair, socially efficient, and less brittle against adversarial manipulation.
Objectives: The objective of this PhD project is to develop a dynamic game-theoretic framework to improve the safety and trustworthiness in AI systems. Particular emphasis will be placed on systems with complex decision making processes in which humans repeatedly interact with AI algorithms such as virtual economies and cryptoeconomic mechanisms. Based on this framework, the main goal of the project will be to design mechanisms and protocols that will be robust to uncertain and adversarial environments. The proposed mechanisms will be used to identify potential critical (or phase) transitions that can occur in these systems, and to develop strategies for predicting and mitigating such events. This will involve analyzing the impact of different game-theoretic strategies on stability and robustness, and identifying optimal strategies for achieving safe and trustworthy behavior.
Techniques: The work in this PhD will employ a combination of theoretical work in the above areas and empirical evaluation of the proposed mechanisms or algorithms using simulations and real-world datasets. The approach will integrate formal mathematical methods with game dynamics and machine learning techniques (e.g., offline, or online reinforcement learning) to model and analyze the behavior of the considered AI systems over time.
Impact: The potential impact of this project is significant and manifold. By improving the safety and trust in AI and cryptoeconomic systems, this project has the potential to enable the wider adoption of these technologies, and to unlock their full potential for a wide range of applications. This could have significant benefits for society, including improving the efficiency and reliability of critical infrastructure, and enabling the creation of innovative and reliable services.