|M.Sc Student||Sarah Keren|
|Subject||Tutoring as Sequential Decision Making|
|Department||Department of Industrial Engineering and Management||Supervisor||Full Professor Domshlak Carmel|
|Full Thesis text|
Automated tutoring systems try to find a tutoring plan that is as tailored as possible to the conditions of the specific tutoring session. We are interested in tutoring scenarios in which the system receives as input a set of limited resources, a set of available tutoring actions with stochastic effects, a set of items to be tutored and a precedence relation between the items, which allows the tutoring of an item only after all its predecessors have been learned. The goal is to find, given the constraints, an optimal tutoring strategy that maximizes the expected reward collected within a tutoring session.
Generally, the job of the system can be divided into two main tasks: the first being the selection of the next item to teach at each stage of execution and the second being the selection of the best tutoring policy for the chosen item. This separation naturally induces a hierarchical structure that can be exploited in the creation of efficient search strategies. Accordingly, we formulate the Stochastic Tutoring Planning (STP) model as a generalization of the Stochastic Over Subscription Planning Problem (SOSP). The SOSP is a hierarchical Markov Decision Process (MDP), in which the root process is responsible of selecting a sub-process to execute, and each sub-process tries to maximize the accumulated reward until it propagates control back to the root-process. The compilation process and the required adjustments are presented in our work. We show that the result allows us to exploit the informed and uninformed algorithms developed for the SOSP domain almost without modification. We concentrate our effort on developing efficient heuristics estimations for the hierarchical informed search which proved to be efficient not only for the STP domain but to other hierarchical domains, including the Rover domain, which inspired the creation of the SOSP model.