|
Projects
The purpose of this set of projects is to apply statistical methods in modeling and optimizing different aspects of human behavior and human group behavior.
| Latent Structure Influence Model |
Wen Dong |
|
The latent structure influence model is a state-space model that effectively compresses the large latent-state space by exploring and exploiting the structure in this space. Compared with other state-space models that do not consider the structure, the latent structure influence process captures more information, is more frugal in its number of parameters, and is less likely to overfit. We have used the latent structure influence process to model the interaction of over one hundred persons and got interesting structural information as well as summaries of individual behaviors that could not be found otherwise. |
| Correlated, restless, multi-agent and multi-armed bandits |
Ankur Mani |
|
We study two aspects of variation to the classical multi-armed bandit problem. First, restless bandit problem considers m out of n projects being operated at any time. It defreezes the unattended projects and allows them to evolve over time. They may also contribute reward as they change state. Whittle (1988) showed a special type of index, regarded as subsidy index, can be assigned to each project. He also conjectured the index policy which operates the m projects of the greatest index is asymptotically optimal as m and n go infinity. This conjecture was proved to be false generally by Weber and Weiss (1990) though it breaks under rare conditions. They derived a sufficient condition for the index policy to be valid. Second, introducing more than one player into the problem gives rise to the multi-agent characteristics. In this version, it is also interesting to characterize equilibrium properties of the optimal policy. We formulate a simple model involving multiple player and multi-armed bandit. The problem structure is laid out and a few properties presented in the form of lemmas. Our work is an early stage attempt in this field.
In future, we propose to fully characterize the equilibrium and the optimal policies for each player and the optimal policy for the team play. We also want to investigate the scenario when the bandits are not independent. This additional complexity makes the problem more challenging but is often seen in real life, for example consider several health agencies working to control epidemic in a social network. The people are bandits and their health states are the states of the bandits. The disease can spread from one person in the network to another and the actions of the agencies can cure the disease and prevent the spread. Similar scenarios can be seen in marketing and advertisement. |
| Social Network from Mobile Phone Usages |
Nathan Eagle |
|
The cellphone network experiment gives quantitative analysis of the periodicities of human behavior and human group behavior, as well as the friendship/workgroup relationships. The analysis is based on one year worth of cellphone related data with 81 MIT Media Laboratory students and Sloan Business School student, the largest and most comprehensive data set by our knowledge when the data set was published. |
| Learning Humans |
|
|
This repository contains a large collection of our earlier projects in applying computer vision, speech recognition, statistical learning, and graphical model algorithms to the understanding of human interaction, human behavior, and human environment. |
Recent Publications
Wen Dong, Alex Pentland, Bruno Lepri, Fabio Pianesi, Alessandro Cappelletti, and Massimo Zancanaro, Using the Influence Model to Recognize Functional Roles in Meetings, Ninth Int'l Conf on Multimodel Interfaces, Nov 12-15, 2007, Nagoya Japan.
Wen Dong, (Advisor: A. Pentland), Influence Modeling of Complex Stochastic Processes, Masters Thesis in Media Arts and Sciences.
Nathan Eagle and Alex Pentland, Eigenbehaviors: Identifying Structure in Routine, Ubicomp '06, September 17-21, 2006, Orange County, CA.
Wen Dong and Alex Pentland, Multi-sensor Data Fusion Using the Influence Model (April 2006), Body Sensor Networks Workshop, April 2006, Boston, MA
|
