Scott C. Proper
346 NW 15th, Apt. C
Corvallis, OR 97330
phone: (541)754-7952
e-mail: proper@eecs.oregonstate.edu
homepage: http://web.engr.oregonstate.edu/~proper/


Objective

  • To contribute to the computer science community via research in artificial intelligence. Specifically, I plan to finish my PhD and go on to a career in research, continuing work in reinforcement learning. I am currently seeking a post-doc position or an industry research position.

  • Education

  • B.S. in Computer Science.
  • Montana State University, Bozeman

  • PhD in Computer Science
  • Oregon State University
    Expected graduation date: Fall 2009

    Computer related skills

  • Programming languages: C/C++, BASIC, Lisp, Assembly, Perl, SQL
  • Platforms: Linux, Windows, UNIX
  • Web Skills and languages: HTML, Java, CGI
  • Relevant coursework

  • Software Engineering, Programming Language Design
  • Theory of Computation, Operating Systems, Compilers
  • Computer Architecture, Computer Networks
  • Computer Graphics, Image Processing, Databases
  • Artificial Intelligence, Reinforcement Learning, Cybernetics
  • Algorithms and Data Structures, Bayesian Networks, Graph Theory
  • Work History

  • Webmaster, Jet Propulsion Laboratories, May 1999-December 1999.
  • Primary designer of a web page for engineers, served across the JPL intranet.

  • Programmer, Dyonjet Research, 2001
  • Primary programmer of numerous in-house utilities. Designed GUI software for release to the public.

  • Research Assistant, Oregon State University, 2002-Present
  • Current work includes research on scaling reinforcement learning, particularly in multiagent domains.

    Research Contributions

  • Developed several techniques to mitigate the "three curses of dimensionality" (explosions in state, action, and stochastic branching factor) of reinforcement learning problems, including a new kind of function approximation that generalizes both tables and linear functions -- "Tabular Linear Functions" -- and the creation of a new average-reward model-based value iteration algorithm based on afterstates, "ASH-learning".
  • Developed a new technique -- "Assignment-based Decomposition" -- for decomposing states and actions in multi-agent, multi-task domains that greatly mitigates the three curses of dimensionality by dividing the action-selection step of a reinforcement learning algorithm into two stages: an upper assignment level and a lower task performance level.
  • Further developed and expanded upon assignment-based decomposition by showing how to integrate coordination graphs into the lower task performance level, how to use it together with transfer learning to enable multi-agent domains to scale to large numbers of agents, and how to use search techniques to quickly assign agents to appropriate tasks.

    Publications

  • Proper, S., Tadepalli, P., Tang, H., Logendran, R., A Reinforcement Learning Approach for Product Delivery by Multiple Vehicles,
    for IIE/IERC 2003: Institute of Industrial Engineers/Industrial Engineering Research Conference.

  • Proper, S., Tadepalli, P., Scaling Average-reward Reinforcement Learning for Product Delivery,
    for AAAI Real Life Reinforcement Learning Fall Symposium 2004.

  • Proper, S., Tadepalli, P., Scaling Model-Based Average-reward Reinforcement Learning for Product Delivery,
    in ECML 2006: Proceedings of the 17th European Conference on Machine Learning, p 735-742.

  • Proper, S., Tadepalli, P., Solving Multiagent Assignment Markov Decision Processes,
    in AAMAS 2009: Proceedings of the 8th International Joint Conference on Autonomous Agents and Multiagent Systems., p 681-688

  • Proper, S., Tadepalli, P., Transfer Learning via Relational Templates,
    in ILP 2009: Proceedings of the 19th International Joint Conference on Inductive Logic Programming. (to appear)