Name:
Joseph O'Sullivan

Degrees Held:
BA(Mod) Computer Science (Trinity College, Dublin, 1992)
MSc (CMU, 1997)

Year Entered Ph.D. Program:
1992

Advisor:
Tom Mitchell

Research Interests:
Life-Long Learning, Action Models, Learning, Machine Learning, Mobile Robotics, Perception, Real World Experimentation, Reinforcement Learning, Xavier, Amelia

Description of Research:
My primary work is in the Learning Robot Laboratory, attempting to develop robots which learn from experience and in the real world. Initially I focussed on reinforcement learning: an inductive method for learning control strategies based on delayed reward. A tangible contribution of this work was developing one of the first applications of ``reinforcement learning'' from vision in a real world complex domain (that is, to fixate and approach a cup). While seeing that such learning techniques could be successful in acquiring control strategies for simple robot tasks, currently we are researching learning techniques which will scale up to more complex and realistic tasks.

This first step in this research was studying Explanation-based Neural Network Learning . This technique unifies neural network and explanation-based learning approaches, combining the advantages of inductive and analytical learning techniques. Our robot test-bed wanders the corridors of CMU, inductively learning action models which are then used as a domain theory in analytically learning corridor concepts (that is, hallways, junctions etc) based upon functional definitions. The sensation of the world contains both vision and sonar information, with the fusing of these modalities being part of the learning task.

This, and much more work, has led to my proposing a thesis on "Transfer of Learning Knowledge in Life-Long Learning Agents ".

The robot testbed in the Learning Robot Lab, Xavier, was designed and constructed internally within the laboratory in 1993. As one of the members of this process and of the subsequent software development for/entry into the various AAAI Robotic Competitions, I maintain a practical secondary interest in lowlevel aspects of robotics (operating systems, device drivers, etc) -- all an important consideration for long-life autonomous robots.

Bibliography

    Book Chapters:

  1. Joseph O'Sullivan, Tom M. Mitchell and Sebastian Thrun. ``Explanation-Based Learning from Mobile Robot Perception''. In Symbolic Visual Learning, Editors: K. Ikeuchi & M. Veloso. Oxford University Press. 1997. [Abstract] [Postscript]

  2. Sebastian Thrun and Joseph O'Sullivan. ``Clustering Learning Tasks and the Selective Cross-Task Transfer of Knowledge'' in S. Thrun and L.Y. Pratt (eds), Learning To Learn. Chapter 10. Kluwer Academc Publisher. 1997

  3. Mark T. Keane, Barry Smyth and Joseph O'Sullivan. ``Dynamic Similarity: A Processing Perspective on Similarity'' in Ramscar, Hahn, & Pain (eds), Similarity And Categorisation,. Cambridge University Press. forthcoming: Summer, 2000.

    Refereed Papers:

  4. Joseph O'Sullivan, John Langford, Rich Caruana and Avrim Blum. ``FeatureBoost: A Meta Algorithm for Unabridged Learning that Improves Model Robustness''. To appear in International Conference of Machine Learning 2000.

  5. Reid Simmons, Joaquin Fernandez, Rich Goodwin, Sven Koenig and Joseph O'Sullivan. ``Xavier: An Autonomous Mobile Robot on the Web'', To appear in IEEE Robotics and Automation Magazine, 2000.

  6. Boris Gutkin, G. Bard Ermentrout and Joseph O'Sullivan. ``Layer 3 Patchy Recurrent Excitatory Connections May Determine The Spatial Organization of Sustained Activity in the Primate Prefrontal Cortex''. In Neurocomputing, Elsevier Science, forthcoming: Summer 2000

  7. Joseph O'Sullivan. ``Transferring Learned Knowledge in a Lifelong Learning Mobile Robot Agent'' In European Workshop on Learning Robots. Edinburgh, UK. July, 1998.

  8. Rich Caruana and Joseph O'Sullivan. ``Multitask Pattern Recognition for Autonomous Robots'' IEEE/RSJ International Conference on Intelligent Robotic Systems (IROS 98), Victoria, B.C., Canada. October, 1998.

  9. Rich Caruana and Joseph O'Sullivan. ``Multitask Learning for Vision-Based Autonomous Robots'' International Conference on Artificial Neural Network (ICANN 98). Skövde, Sweden. September, 1998.

  10. Reid G. Simmons, Richard Goodwin, Karen Zita Haigh, Sven Koenig, Joseph O'Sullivan, Manuela M. Veloso. "Xavier: Experience with a Layered Robot Architecture". In "Intelligence", edited by Manuela M. Veloso. (This paper is an extended version of the Agents'97 paper below). [Postscript]

  11. Reid Simmons, Richard Goodwin, Karen Zita Haigh, Sven Koenig, Joseph O'Sullivan. ``A Modular Architecture for Office Delivery Robots''. In Autonomous Agents 1997, February 1997. [Abstract]

  12. Sebastian Thrun and Joseph O'Sullivan. ``Discovering Structure in Multiple Learning Tasks: The TC Algorithm''. In Proceedings of ICML-96, July 96, Torino, Italy. [Abstract] [Postscript]

  13. R. Simmons, S. Thrun, G. Armstrong, R. Goodwin, K. Haigh, S. Koenig, S. Mahamud, D. Nikovski, and J. O'Sullivan. ``Amelia'' (abstract). In Proceedings of the AAAI Thirteenth National Conference on Artificial Intelligence, Portland, Oregon. 1996

  14. Joseph O'Sullivan.``The CMU Learning Robot Laboratory Data Toolkit'' In Proceedings of the MLC-Colt '94 Workshop on Robot Learning. Rutgers, The State University of New Jersey, New Brunswick, July 9, 1994. [Abstract] [Postscript]

  15. Tom M. Mitchell, Joseph O'Sullivan and Sebastian Thrun. ``Explanation-Based Learning for Mobile Robot Perception''. In Proceedings of the MLC-Colt '94 Workshop on Robot Learning . Rutgers, The State University of New Jersey, New Brunswick, July 9, 1994. [Abstract] [Postscript]

  16. Joseph O'Sullivan. ``Towards a Robot Learning Architecture''. In Learning Action Models - Papers from the 1993 AAAI Workshop, Technical Report WS-93-06, AAAI Press 1993, ed. Wei-Min Shen, pages 47-51. [Abstract] [Postscript]

    Unrefereed Papers:

  17. Sridhar Mahadevan, Andrew Moore and Joseph O'Sullivan, `` Report on Reinforcement Learning and Machine Learning for Manufacturing.'' in Automated Learning and Discovery: State-Of-The-Art and Research Topics in a Rapidly Growing Field, edited by Sebastian Thrun, Christos Faloutsos, Tom Mitchell, Larry Wasserman. CMU-CALD-98-100, Center for Automated Learning and Discovery, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, September 1998. [Postscript]

  18. Sebastian Thrun and Joseph O'Sullivan. ``Learning More from Less Data: Experiments with Lifelong Robot Learning''. Seminar Digest, IEE, London, 96. [Abstract] [Postscript]

  19. Joseph O'Sullivan.``Transfer of Learned Knowledge in Life-Long Learning Agents''. CMU Thesis Proposal, February 1997. [Abstract] [Postscript]

  20. Joseph O'Sullivan. ``On Similarity and Neural Networks''. Trinity College Dublin, Bachelor's Thesis, June 1992.

    Technical Reports:

  21. Joseph O'Sullivan, Karen Haigh and Greg Armstrong. ``Xavier - the Manual''. Carnegie Mellon University Learning Robot Lab Internal Document. April 1997 [Abstract] [Postscript]

  22. Sebastian Thrun and Joseph O'Sullivan. ``Clustering Learning Tasks and the Selective Cross-Task Transfer of Knowledge''. Technical Report CMU-CS-95-209, Carnegie Mellon University, Pittsburgh, PA, 1995. [Abstract] [Postscript]

  23. Joseph O'Sullivan and Mark T. Keane. ``Using Neural Network Techniques to Model Similarity Phenomena''. Trinity College Dublin, Technical Report TCD-CS-92-36, October 1992. [Abstract] [Postscript]

  24. John Mitchell and Joseph O'Sullivan. ``N-Net: An Exploratory Tool for the Investigation of Feed-forward Neural Networks as Function Approximators'', Hitachi Dublin Laboratory, Technical Report HDL-TR-91-029, October 1991. (This appeared in ``Mitchell, J., (1992) Geometrical Interpretation of Hidden Layer Units in Feedforward Neural Networks Network Computation in Neural Systems Vol 3 No 1 February 1992'').

    Unpublished:

  25. Joseph O'Sullivan. `` Integrating Initialization Bias and Search Bias in Artifical Neural Networks''. 1996 [Abstract] [Postscript]

  26. Joseph O'Sullivan and Sebastian Thrun. ``A Robot That Improves Its Ability To Learn.'' Unpublished. 1996 [Abstract] [Postscript]

  27. Rich Caruana & Joseph O'Sullivan. ``Backprop Nets Do Unsupervised Clustering of Outputs''. 1998

  28. Joseph O'Sullivan. ``Sequential Multitask Learning for a Lifelong Learning Agent''. 1998

Joseph Kieran O'Sullivan
Last modified: Wed Dec 27 04:41:11 EST 2000