by
Abstract:
This thesis addresses the challenge of peer prediction, which seeks to elicit private information from rational agents without the requirement that ground truth is eventually revealed. The classical peer prediction method provides a solution to the peer prediction challenge. It compares the reported information of an agent with the reported information of another agent, and computes a payment rule that implements truth revelation in a strategic equilibrium. However, the algorithm computing the payments critically depends on the method's assumption that all agents share the same prior beliefs and that the algorithm ("mechanism") knows these beliefs. In this thesis, I relax this common knowledge assumption. I first design the "Robust Bayesian Truth Serum" (RBTS), which asks agents for two types of reports, the report of the private information it is interested in and a prediction report corresponding to an agent's belief about the private information of other agents. RBTS dispenses with the assumption that the agents' prior beliefs need to be known to compute the payments. It does, however, still rely on the agents sharing the same prior beliefs. My second contribution is the design of "subjective-prior peer prediction mechanisms," which further reduce the assumption of common knowledge. As in RBTS, they do not require knowledge of the agents' prior beliefs. Moreover, they allow the agents' prior beliefs to be subjective, i.e. different from one another. My third contribution is the study of "effort-incentivizing peer prediction." In many applications of interest, the information that seeks to be elicited first needs to be acquired. When this information acquisition requires costly effort, agents may have an incentive to avoid it and choose to guess instead. Addressing this problem, I suggest payments, where only agents with good enough information have an incentive to participate in the mechanism. Agents not investing effort and agents with low quality, choose to pass, effectively self-selecting according to quality.
Reference:
Robust Peer Prediction Mechanisms J. WitkowskiPhD thesis, Department of Computer Science, Albert-Ludwigs-Universit├Ąt Freiburg, 2014
Bibtex Entry:
@PHDTHESIS{witkowski:2014,
  author =	 {Witkowski, Jens},
  title =	 {{Robust Peer Prediction Mechanisms}},
  school =	 {Department of Computer Science,
                  Albert-Ludwigs-Universit\"{a}t Freiburg},
  year =	 {2014},
  month = 	 {May},
}