Email: gfarina AT cs.cmu.edu

Gabriele Farina

I am currently a Research Scientist at FAIR (Meta AI).

I will join MIT EECS as an Assistant Professor in Fall 2023.

Before that, I spent six years as Ph.D. student in the Computer Science Department at Carnegie Mellon University, where I was advised by Tuomas Sandholm and was part of the Electronic Marketplaces Lab. I was supported by a 2019-2020 Facebook Fellowship in the area of Economics and Computation.

Before CMU, I was an undergraduate student at Politecnico di Milano, where I was advised by Nicola Gatti.

Curriculum Vitae Publications

Computational Game Solving (CMU 15-888, F21)


Research Interests

With my research, I seek to provide solid theoretical and algorithmic foundations for computational, strategic (i.e., game-theoretic) decision-making under imperfect information. To achieve that, I combine and advance techniques and notions of strategicness from game theory together with modern tools from machine learning (especially online learning), optimization, and statistics.

Without a game-theoretic understanding of strategicness, machine learning techniques alone are incapable of understanding and responding to deception from strategic agents, as well as safely reason about imperfect information in adversarial settings. (For example, methods from single-agent reinforcement learning are hardly applicable, given that the learning of the other (adversarial) agents makes the environment nonstationary.) On the other hand, online learning, statistical, and optimization tools provide ways to operationalize static ideas of game-theoretic optimality, giving efficient means of optimizing strategies, learning opponent models, and predicting outcomes.

No-Regret Learning Dynamics

No-regret learning dynamics for games are a fascinating theoretical problem ("can local learning result in global game-theoretic equilibrium?"), as well as currently the most practically scalable technique we know for training strong agents in large games.

Publications on this topic: [ = recommended read]

NeurIPS 2022
pdf
NeurIPS 2022
pdf
ICML 2022
pdf
ICML 2022
pdf
STOC 2022
pdf
AAAI 2022
pdf
EC 2021
pdf
AAAI 2021
pdf
AAAI 2021
pdf
AAAI 2021
pdf
ICML 2020
pdf
NeurIPS 2019
pdf
ICML 2019
pdf
ICML 2019
pdf
AAAI 2019
pdf
NeurIPS 2018
pdf

Correlated and Coarse Correlated Equilibria

Most of the literature on strategic decision-making so far has focused on the task of computing optimal strategies for individual agents that seek to maximize their own utility. On the other hand, many realistic interactions require studying correlated strategies. The study of the geometric and analytical properties of correlated strategies in extensive-form strategic interactions is a fundamental question with applications to diverse settings, and is yet to be fully explored.

Publications on this topic: [ = recommended read]

NeurIPS 2022
pdf
EC 2022
pdf
EC 2022
pdf
STOC 2022
pdf
NeurIPS 2020
pdf
NeurIPS 2020
pdf
AAAI 2020
pdf
NeurIPS 2019
pdf

Team Games and Team Equilibria

Most of the literature on strategic decision-making so far has focused on the task of computing optimal strategies for individual agents that seek to maximize their own utility. On the other hand, many realistic interactions require studying correlated strategies. The study of the geometric and analytical properties of correlated strategies in extensive-form strategic interactions is a fundamental question with applications to diverse settings, and is yet to be fully explored.

Publications on this topic: [ = recommended read]

NeurIPS 2022
pdf
ICML 2021
pdf
NeurIPS 2018
pdf

Human Modeling, Robustness to Mistakes, and Equilibrium Perfection

Nash equilibrium is the most seminal solution concept in game theory. However, in some strategic interaction it is too restrictive, assuming that the agents have unlimited computational power to come up with the optimal solution to the interactions. On the other hand, in other strategic interactions it is too permissive, prescribing unsatisfactory strategies. For example, in the case of multi-step interactions, one limitation is that some Nash equilibria do not prescribe optimal play after the player or the opponent has made a mistake (sequential irrationality). Resolving these issues is important to develop algorithms that are ready to operate in the real world.

Publications on this topic: [ = recommended read]

ICML 2022
pdf
NeurIPS 2021
pdf
AAAI 2019
pdf
NeurIPS 2018
pdf
IJCAI 2018
pdf
AAAI 2018
pdf
ICML 2017
pdf
IJCAI 2017
pdf
AAAI 2017
pdf