Third call for collaborative research proposals

Winners of third CIVICA Research call announced

CIVICA is the European University of Social Sciences. It unites ten leading European higher education institutions in the social sciences, humanities, business management and public policy. It regularly offers research grants to its partner universities, supporting travel for joint work and workshops.

As part of the CIVICA priority area "Data Driven Technologies for Social Sciences", the project "Multi-agent learning and equilibrium" will run October 2022-October 2023 in London, Milan and Stockholm. It is a cross-disciplinary project at the departments of Mathematics (LSE), Computer Science (University of Bocconi), and Economics (SSE - Stockholm School of Economics), based on the expertise of the investigators in game theory, machine learning, and economic equilibrium concepts. At LSE, the researchers are Galit Ashkenazi-Golan, Katerina Papadaki, and Bernhard von Stengel; at Bocconi, Andrea Celli; and at SSE, Mark Voorneveld. A first joint meeting took already place November 9-11 at LSE, where Andrea Celli and Mark Voorneveld gave seminars on their work. PhD students in game theory are also involved, so far Sahar Jahani and Ed Plumb.

Many economic decisions are made automatically by learning algorithms, for example the dynamic pricing of hotel rooms. Can these algorithms learn to *collude*, with higher prices than what would be expected from perfect competition? Some evidence suggests this may happen. With a mathematical model of a "pricing game", this is the first question studied in this project. The more general question is to understand better how machine learning works when multiple independent agents learn on their own. The research challenges are much more complex that in single-agent learning, which has been very successful in, say, image recognition. Learning during the interaction among multiple agents may lead to unusual optimizing behaviour against each other. They may be stuck in a "bad equilibrium" instead of cooperating. On the other hand, one may not want hotels to "cooperate" (that is, collude) with higher prices. All this needs to be better understood.

The project will develop a novel modular approach with exchangeable "parts" for the given game of competition and cooperation, the machine learning method, and its evaluation in various "equilibrium" scenarios of reasonable behaviour of the agents. These parts can be varied separately, within a new open-source software framework. 

You can read more about CIVICA's Research call here.