Multi-agent optimization with privacy protection

  • Date April 12, 2018
  • Hour 3pm
  • Room GSSI Main Lecture Hall
  • Speaker Fabian Wirth (University of Passau)


We consider a relatively simple class of distributed optimization problems, in which a number of agents want to reach the social optimum of consumption of a common limited resource. The problem is that the agents are unwilling to reveal their personal cost function for using the resource, nor do they want to reveal their individual consumption. A central entity is able to measure overall consumption of the resource and to signal whether overall consumption requests exceed the available resource. Using a probabilistic response of the agents exclusively based on their private information, i.e. cost function and past consumption, it is possible to describe an iterative scheme that converges almost surely to the optimal point of this multiagent optimization problem.
The proposed algorithm is based on the AIMD algorithm, the evolution of which can be described by products of matrices from a given set. The algorithm can be described in terms of a non-homogeneous Markov chain. Paracontractivity properties of the matrices ensure the desired convergence of the long-term averages of consumption.