|Ph.D Student||Davidovitch Lior|
|Subject||Strategic Interactions under Severe Uncertainty|
|Department||Department of Mechanical Engineering||Supervisor||PROFESSOR EMERITUS Yakov Ben-Haim|
|Full Thesis text|
Decision making is often made in an environment of uncertainty. This research focuses on Knightian uncertainty, uncertainty which cannot be reliably expressed probabilistically. We examine how to choose a strategy when confronted with an interactive adversarial agent: an agent reacting perniciously to the decision maker's choice.
This interaction tends to increase the uncertainty. The decision maker often cannot know precisely the goals and preferences of the adversarial agent. This lack of knowledge can rarely be reliably expressed probabilistically.
We examine the use of info-gap decision theory as a decision supporting tool in face of Knightian uncertainty. Info-gap decision theory urges the decision maker to satisfice his criteria rather than to optimize them. That is, instead of finding the best alternative under our unreliable best model of the world, find an alternative which is good enough for a wide set of contingencies, or robust. We must do this in the absence of knowledge of a worst case.
We demonstrate the possible contribution of info-gap analysis to situations of uncertain strategic interaction by examining three problems: profiling, voting, and detection. We also study a theoretical relation between robustness and probability.
Profiling is an actuarial method of law-enforcement, based on the fact that different groups within the population have different responsivenesses (or elasticities) to policing. These differences allow a decision maker to divert policing resources to highly responsive groups at the expense of non-responsive groups, thus decreasing the total crime rate.
Elections are generally seen as a method for reflecting the preferences of the voters. However, a voter may have an incentive to strategically mis-represent his preference, based on the expected voting patterns of other voters. This kind of behavior is called strategic voting.
Designing a surveillance system for the detection of an infiltrating agent is often done in an uncertain environment, even if the agent is not assumed to respond to the design of the system. When incursion attempts are very rare, it is difficult to predict the characteristics of the next attempt, or even to estimate the detection capabilities of system components.
In all the above problems, a decision maker must make his choice under severe uncertainty, of responsiveness functions, preferences of other voters, or characteristics of incursion attempts.
Finally, we examine the connection between robustness and probability of success. We will discuss the conditions under which robustness may be used as a proxy for the probability of success.