I am currently working on developing an automated negotiation agent at CMU. I had the opportunity to work on numerous diverse but interrelated topics during the course of my PhD and Postdoc.
1. Automated Negotiation: Negotiation is an interactive process by which multiple self-interested parties with limited common or uncertain knowledge about opponents try to arrive at an agreement over a set of issues with possibly conflicting preferences over the issues. Most work in computational modeling to-date has focused on the outcome of a negotiation. My focus is to develop a general purpose negotiation agent with special emphasis on capturing the process of negotiation and not just the outcome. To capture the evolution of the negotiation process, I modeled the negotiating agent as a dynamical system that evolves in time. In particular, the negotiation problem is modeled using a decision-theoretic framework namely Partially Observable Markov Decision Process (POMDP) ({See MICON 09}).
2. Security using Game Theory: Security is defined as the ability of the system to deal with intentional threats from other agents/systems. My basic assumption is that the agent/agent-team knows the adversaries actions/payoffs but their type is unknown and hence modeled as a Bayesian Game. I developed a heuristic algorithm named ASAP{See AAMAS 07} and an exact algorithm named DOBSS{See AAMAS08} to solve this security game. DOBSS became heart of the ARMOR system currently being tested for security scheduling at the Los Angeles International Airport.
3. Security by Policy Randomization: In this project, I developed security algorithms where the agents have no model of the adversary. We therefore play a information minimization game with the adversary where we do intentional randomization of agents policies in single/decentralized (PO)MDPs without significantly sacrificing rewards or breaking down coordination. Following are my contributions in this direction {See AAMAS 06}:
- Three novel algorithms, one based on NLP (Non Linear Program) and two based on LPs to randomize single agent policies while attaining certain level of expected reward.
- RDR (Rolling Down Randomization algorithm) to efficiently generate randomized policies for Decentralized POMDPs.
4. Teamwork with resource constraints: I introduced EMTDP, a distributed MDP framework where agents not only maximize expected team reward but bound the expected resource consumption {See AAMAS 04}. Following is an overview of this work:
- Resource constraints in team settings lead to randomized policies.
- Randomized policies in team settings lead to miscoordination.
- EMTDP tranformation avoids miscordination even if randomized policies in team settings.
- Developed nonlinear programming techniques as solution method for the new framework.
5. Self vs Team Interest in Multiagent Systems: This project was based on the Electric Elves domain, a multiagent meeting scheduling system implemented at USC. I applied my initial formalization of the notion of self interest (competition) vs team interest (collaboration) for this domain. For more details see GTDT 03.
During my undergraduation I also worked on building a realistic Traffic Simulator using the BDI architecture {See AAMAS 02}.
These projects provided me an oppurtunity to program in C, C++, Java, Perl and work with softwares like Lingo, Ampl, Gambit toolkit etc.