Decision-support

IDEM - Information and Decision Making

Date:
Changed on 14/10/2021
IDEM aims at characterizing the interaction between data acquisition and information processing in decentralized decision making by bringing together tools from information theory and game theory. This characterization is central to the understanding of certain problems, such as decentralized optimization and machine learning, that are subject to local information constraints.

What does "IDEM" mean and what are your main research themes?

The exploratory action IDEM (information and decision making) seeks to develop mathematical models to predict the behavior of autonomous agents interacting in decentralized decision making.
 
These models allow analyzing machine learning problems subject to local information constraints. To this extent, IDEM focuses on identifying fundamental limits on key performance metrics. Fundamental limits, as the laws of Physics, are those bounds that technological advancement cannot overcome. Hence, these limits establish a benchmark with which new ideas, developments and algorithms must be compared to for determining their optimality. For instance, in the  context of supervised learning, a relevant  performance metric is the generalization error, which assesses the capacity of a machine to handle a situation for which it has not been trained.  
 
By modelling these fundamental limits, it will be possible, for instance, to determine the minimum amount of data required by a machine learning system to successfully handle unexpected or rare events. 

Is it more of a basic or applied research topic?

IDEM undertakes fundamental research in applied mathematics at the intersection of information theory and game theory.

The joint use of these theories allows to quantitatively model agents that possesses different amounts of information. Interestingly, the notion of information in the sense of Shannon is rarely taken into account in classical game formulations. Neither are the intricacies of physical channels through which information is sent, e.g., noise, interference, etc. These new considerations parametrize most of the concepts in game theory providing new elements for the analysis of decentralized decision making. These theoretical results have immediate applications in the context of federated learning, particularly on practical policies with which distributed learning machines shall reveal or hide information to their peers in order to improve global metrics, for instance, average generalization errors.

Within this context, the long term objective is to traverse the traditional boundaries between the fields of information theory, data sciences and game theory.

How is the project exploratory?

The exploratory nature of IDEM lies on the fact that it brings together tools from information theory, data sciences, and game theory to study the impact of information on decentralized decision making.

There exist preliminary works involving these theories, but the simultaneous consideration of game and information theories to study decision making models is at this moment a largely uncharted territory. 

Do you have academic or industrial partners?

This exploratory action is developed within the existing cooperation of the team NEO with scholars at the University of Sheffield, the National Chiao Tung University in Taiwan and Princeton Universit.

Contacts

Samir Perlaza, équipe-projet Neo

Samir Perlaza

Alain Jean-Marie, équipe-projet Neo

Alain Jean-Marie