I was a researcher with the TeamCore Research Group led by Professor Milind Tambe on the USC Landroids (now the MAAF project) Project which seeks to apply the Distributed Constraint Optimization Problem (DCOP) Framework to a robotic sensor network. DCOP is a general purpose algorithmic framework for multi-agent coordination previously used for agent-based scheduling and other domains. Our mission with this project was see what kind of results this agent coordination framework can give us on a robotic sensor network. We saw some good results in simulation and I was excited to see what could happen with real world robots.
In the summer of 2008, I programmed a robotic sensor network of 5 iCreates to prototype the group’s system.
I was able to scale the network up to 8 robots. Soon I hope to have 10 iCreate robots in the network executing our algorithms. DCOP algorithims facilitate coordination amongst the robots as they decide within their local networks which robots should move a couple inches so that the local signal strengths get better. These local optimizations have experimentally been shown to have a beneficial effect on global throughputs and ping rates. The results we’ve gotten indicate that with small scale movements alone the throughput of the robotic network can be improved nearly 50%!
Earlier I worked on decision tree analysis for DCOPs. DCOP algorithms have various tradeoffs and when applying these algorithms to real world problems, these tradeoffs need to be optimized. That’s where machine learning can help us. My work involved utilizing decision trees to mine structure out of the parameter spaces of a robotic configuration. This work was published as part of our group paper at the Distributed Constraint Reasoning (DCR) conference at IJCAI 2009.
I am currently extending this work to develop a new class of Hybrid DCOP algorithms that combine the benefits of prior algorithms but use decision trees to switch between algorithms to get the optimal reward.
Update: Landroids project code is now available via the Robotic DCEE Open Source project. See: https://github.com/prateekt/roboticdcee for details.
- Matthew Taylor, Manish Jain, Prateek Tandon, Milind Tambe. Using DCOPs to Balance Exploration and Exploitation in Time-Critical Domains. Distributed Constraint Reasoning Workshop at IJCAI 2009.
- Marcos A. M. Viera, Matthew E. Taylor, Prateek Tandon, Manish Jain, Ramesh Govindran, Gaurav Sukhatme, Milind Tambe. A Distributed Reasoning Approach to Mobile Mesh Network Optimization.
- Scott Alfeld, Matthew E. Taylor, Prateek Tandon, Milind Tambe. Towards a Theoretic Understanding of DCEE. Distributed Constraint Reasoning Workshop at IJCAI 2010.
My work with the TeamCore Research Group is supported by a Viterbi Merit Research Scholarship and (previously) by a Vice Provost Research Fellowship. Special thanks to the sponsoring entities.