Security in Multiagent Systems by Policy Randomization

  • Implemented algorithms to randomize single agent MDPs by maximizing a weighted entropy function and maintaining a certain threshold of reward.
  • Implemented Rolling Down Randomization (RDR), that efficiently generates randomized policies for decentralized POMDPs via the single agent LP method without significantly sacrificing rewards or breaking down coordination.
  • Based on this paper.
  • Github Source Code Link.