Dynamic Mechanism Design

People Involved :

According to Myerson, (Noble laureate, 2007) mechanism design is reverse engineering of game theory where we induce a game among self interested, intelligent and rational agents with conflicting interests such that at equilibrium, we achieve certain desirable behavior among them. In practice, many times, the agents may be dynamic and the decision pertaining to the present agents who are departing may affect decisions for the agents yet to arrive. Thus, theory of online algorithms and mechanism design needs to be combined intelligently to achieve a desirable performance.

 


Incentive Compatible Machine Learning

People Involved :

In many real world applications, learning algorithms rely on data obtained from agents who may be strategic and manipulate learning algorithm in their favor. Consider a scenario where central/federal govt planning to learn something about financial conditions of the country based on inputs from state govts and build some machine learning model and take decision about finance. The state govts may misreport the data available with them so as to make machine learning algorithm to take decisions in their favor. Thus, there is need to build learning algorithms that are robust to strategic manipulations by the agents.

 


Optimizing Average Precision using Weakly Supervised Data

People Involved :

Optimizing Average Precision  (AP) for learning parameters of max-margin models using weak supervision.