Phone : +33 (0)4 76 61 50 71 Onno.Zoeter@xrce.xerox.com
Keywords : (Bayesian) Machine Learning, Mechanism Design, Partially Observable Markov Decision Processes, Approximate Inference.
I am interested in constructing systems that take optimal actions in an environment that is driven by random events and intentional actions of self-interested agents. In particular I am intrigued by the following observation: Learning models of human behaviour and using them to rank, filter, assign, etc. are fundamentally different problems from their analogs in say biology and medicine: humans will be aware of their impact on the system and adapt their behaviour.
When applying machine learning approaches to human data sources we need to take into account incentives. Recent results from the fields of mechanism design, Bayesian statistics, and optimal control allow us to do so. An interesting example of misaligned incentives can be encountered when one tries to learn how well matched an on-line advertisement is to a particular target audience. A straightforward learning approach in the currently used auctions can lead to severe possibilities of exploitation. A slight change in the machine learning algorithm and auction mechanism can provably align incentives of advertisers and searchers.
Applications of the optimal control work within Xerox are in supply chain management: the layout of networks and replenishment policies for very large multi-item distribution problems with uncertain and non-stationary demand. The incentives of human users play a critical role in applications for (on-line) market places, for instance in the tracking of characteristics of participants and in finding suitable assignment and pricing rules.
