Background

Updated:  01/2010

My research interests as I thought of them before I started the PhD program are captured here. The following are my high-level areas of interest:

  • Reinforcement learning and Bayesian methods
  • Multi-agent learning, with emphasis on AI in games and multi-robot systems
  • Mechanism design and game theory
  • Applications, such as Smart Grid control systems and human-robot interaction

I have been advised by Manuela Veloso, whose research group tackles the scientific and engineering challenges of creating teams of intelligent agents in complex, dynamic, and uncertain environments, in particular adversarial environments.

Focus Area

Updated:  12/2013

I currently focus on applications in computational sustainability, especially as it relates to the Smart Grid. Planning and operating a large and complex digital ecosystem like the Smart Grid requires advances in control systems, support for dynamic pricing, computational techniques for game-theoretic models and mechanism design, distributed multi-agent based models, and decision-support and optimization tools [1].

I have worked on projects that address several of the above challenges. I am also a co-creator of Power TAC, an open-source agent-based simulation environment for Smart Grid research and competition; this paper presents an overview and this paper serves as the game specification for the competition setting.

I recently defended my thesis on Semi-Cooperative Learning in Smart Grid Agents. The committee was chaired by Manuela and includes Tom Mitchell, Steve Smith, and Yoky Matsuoka of Nest Labs.

Publications

Updated:  12/2013

Semi-Cooperative Learning in Smart Grid Agents. P. Reddy. PhD Thesis. CMU-ML-13-114 Carnegie Mellon University, December 2013. [bib]
Negotiated Learning for Smart Grid Agents: Entity Selection based on Dynamic Partially Observable Features. P. Reddy and M. Veloso. Proceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence (AAAI-13), Bellevue, July 2013. [bib]
Power TAC: A competitive economic simulation of the smart grid. W. Ketter, J. Collins and P. Reddy. Journal of Energy Economics, 39 (2013) 262-270. [bib]
The 2013 Power Trading Agent Competition. W. Ketter, J. Collins, P. Reddy and M. De Weerdt. Erasmus Research Institute of Management Report Series ERS-2013-006-LIS, SSRN 2268852, May 2013. [bib]
Factored Models for Multiscale Decision Making in Smart Grid Customers. P. Reddy and M. Veloso. Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence (AAAI-12), Toronto, July 2012. [bib]
RSSI-based Physical Layout Classification and Target Tethering in Mobile Ad-hoc Networks. P. Reddy and M. Veloso. Proceedings of the IEEE International Conference on Intelligent Robots and Systems (IROS-11), San Francisco, September 2011. [bib]
Learned Behaviors of Multiple Autonomous Agents in Smart Grid Markets. P. Reddy and M. Veloso. Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence (AAAI-11), San Francisco, August 2011. [bib]
Strategy Learning for Autonomous Agents in Smart Grid Markets. P. Reddy and M. Veloso. Proceedings of the Twenty-Second International Joint Conference on Artificial Intelligence (IJCAI-11), Barcelona, July 2011. [bib]
The Power Trading Agent Competition. W. Ketter, J. Collins, P. Reddy, C. Flath and M. De Weerdt. Erasmus Research Institute of Management Report Series ERS-2011-027-LIS, SSRN 1975237, May-December 2011. [bib]

Other Work

Updated:  09/2012

Robotic Ad-hoc Sensor Networks

I worked on DARPA's LANdroids program, whose goal was to create autonomous communication and surveillance networks in urban indoor environments using mobile robots. This paper describes our algorithms which allow the robots to reduce uncertainty about their relative locations and tether themselves to humans and other robots. (The project was covered by Gizmodo and Engadget.)

IO-Aware Algorithms

We extended the standard IO-model to the cloud computing scenario where we assume data server nodes and client processing nodes are separated by TCP/IP connections. We implement cache-aware and cache-oblivious algorithms for scan and merge sort and validate the use of the IO-model for performance analysis in this setting.

Astrophysics Data Mining

The goal here was to develop new statistical measures and data mining methods to describe the positions of galaxies in the universe. We used simulated positions of galaxies, generated according to prevalent cosmological theories, and developed multiple large scale sampling-based techniques for comparing subspaces and for estimating critical epsilon, the percolation threshold.

Multimodal Activity Recognition


This project used multimodal first person sensor data (video camera and inertial motion) from the CMU-MMAC database to identify the actions being performed by the person wearing the sensors. This was accomplished using an SVM-KNN combination with additional HMM-based smoothing. This paper provides further details.