what i'm working on

hybrid value iteration for POMDPs

A drawback in several online POMDP algorithms is that they never update the bounds that have been computed offline. As an online agent expands its search tree, it might be worthwhile to compute a backup operation and add another alpha vector to the value function. Hybrid Value Iteration (HYVI) is a hybrid algorithm that occasionally updates its offline-computed bound on the value function to improve the obtained reward while meeting real-time constraints. Shown in the picture you see an agent "A" acting the in the RockSample domain using HYVI.

POMDP algorithms

As first explained to me in CS411-AI1, a part of AI studies the design of autonomous agents that interact with their environments via sensors and actuators. I often visualize this picture as a broad definition of the field, and one can think of a Partially Observable Markov Decision Process (POMDP) as a mathematical description of just that. I have been studying POMDP solving techniques since I joined the MAS group, and this section of the site is an ongoing effort to organize what I have learned so far. More information here.

offline / online combination

One of the many ways in which POMDP algorithms can be classified is into offline and online methods. In this project we try to derive general methodologies to design a hybrid offline / online agent. Our aim is to invest the right amount of computational power offline in order to leverage an online implementation.

collaborative, computer-supported learning of animal communication

Contemporary goals of science education require that students not only learn the facts and laws but instead understand science as a principled process of inquiry. Students need to be part of directing some of the inquiry and also participate in discussions. In this work, we have designed a computer-supported activity that introduces students to a unique example of animal communication. The activity scaffolds the students' discovery of the bee tail-wagging dance by guiding their inquiry process across four phases with the help of goal directed group tasks, discussions, and experimentation. The natural phenomenon of the tail-wagging dance stands as a unique example of communication in the animal kingdom and its discovery is a quintessential model of the scientific methodology. Middle school students will be immersed in a virtual environment that allows them to observe, hypothesize, experiment, and discuss findings connected to the waggle dance. Aside from the mechanism of the dance itself, our activity promotes learning about the scientific process, control of multiple variables, polar coordinate systems, and teamwork.

past projects

a study of password habits assessment methods

A major focus of human-computer interaction and security research has been to strengthen systems providing sensitive services. Although algorithmic advancements are necessary, overall security offered by a system not only depend on the use of cryptography but also on the humans interacting with it. Many researchers have found insecure habits in the creation and maintenance of passwords, and have regarded humans as the weakest link Several studies have made important advances by studying existing systems, users’ attitudes and habits. To avoid human limitations (memory, computation power) researchers have also proposed new interesting authentication systems. Most of these studies evaluate the efficacy of the proposed systems against existing systems on characteristics like password strength, reuse, etc. They often employ one of three methods to capture human habits, namely: Standard Survey, Creation Method and Real password analysis. Survey techniques gage habits by posing several questions to users in relation to their passwords. Creation techniques pose similar questions about a password created on the spot. Real password studies, as the name suggests, analyze actual passwords of users by means of software tools such as agents deployed over the Internet. We present empirical evidence that conventional data gathering techniques have inherent weaknesses and collect incomparable data. We also propose a new method to capture password-related data, called Direct Metrics, that primarily avoids overly intrusive questions. Further, we conclude with several recommendations for studies that aim to gather password-related information from users.