This page is defunct. It exists to help citeseer find some "relevant" citations.

For a complete listing of Ric Crabbe's Publications, go to his Bio.

Carlos Antonio Acosta Calderon and Huosheng HuRequirements for Getting a Robot to Grow up In the proceedings of 7th European Conference on Artificial Life (ECAL 2003), Dortmund, Germany, September 14-17, 2003

P. Gómez, A. Alvarez, R. Martínez, M. Pérez, V. Rodellar and V. Nieto A DSP-based modular architecture for Noise Cancellation and Speech Recognition. 1998 IEEE Int. Symposium on Circuits and Systems, ISCAS'98 Monterey, California, USA, 31-May/3-June, 1998

Jens Wawerla, Gaurav S. Sukhatme, Maja J. Mataric, Collective Construction with Multiple Robots Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2002) October 4, 2002.

M. McTear, Spoken dialogue technology: enabling the conversational user interface ACM Computing Surveys (CSUR), 34:1, 2002.

Bill Manaris, Natural Language Processing: A Human-Computer Interaction Perspective, In Advances in Computers (Marvin V. Zelkowitz, ed.), vol. 47, pp. 1-66, Academic Press, New York, 1998.

C.A. Acosta Calderon and H. Hu, Robotics Societies: Elements of Learning by Imitation, Proceedings of the 21st IASTED Int. Conference on Applied Informatics, pp. 315-320, Innsbruck, Austria, 10-13 February 2003

Jeanne Fromer, Learning Optimal Discourse Strategies in a Spoken Dialogue Systems (1998) (unpublished S.M. thesis, Massachusetts Institute of Technology) (on file with the Massachusetts Institute of Technology library).

Stoness, Scott C. Continuous Understanding: A First Look at CAFE, Area Paper for URCS, May 2001

Niels Ole Bernsen, Hans Dybkjær and Laila Dybkjær: Elements of Speech Interaction. In Dybkjær, L. (Ed.): Proceedings of the Third Spoken Language Dialogue and Discourse Workshop, Vienna, September 1997.

Introduction

We are developing agents that can build structures out of building materials found in a two dimensional environment consisting of circular objects called discs. Primarily we focus on nests, which are structures that can protect the agents from predators or other dangers like weather. Nests include briar patches and enclosures. Briar patches are collections of material packed close together, preventing passage of predators but not close enough to prevent passage of the agents, and enclosures are surrounding walls with one or more openings. The agents, controlled by artificial neural networks, must complete the task while performing other tasks necessary to their survival. As the agents build the object, they must still be able satisfy their more immediate goals, such as eating, or fleeing predators.

Current research focuses on learning behaviors and building maps of the environment. The agents use a new type reinforcement learning to build sparesely connected neural networks. The also can learn vicariously, that is learn about their environment by observing the results of actions done by peers. Mapping is currently perfromed with concentric rings of neurons that activate

The Environment

The experiments are run in DiscoTech, a continuous two-dimensional world. Everything in the environment (both objects and agents) are circular discs. Each disc is a single color: agents are black, while the passive discs consist of green-colored food discs, blue-colored water discs and building discs of a variety of other colors. Food and water discs can be consumed by the agents. Food discs and water disappear when eaten or drunk.

Agent effectors consist of: turning left or right with some movement forward (i.e. an agent cannot spin in place); moving forward, left, right or backwards; eating; drinking; painting discs (i.e. changing the color of a non-food or non-water disc); and grasping. Grasping must be done continuously over a period of time for an agent to hold onto a disc for that period (Dropping is done as a cessation of grasping).

Time is divided into discrete time-steps. At each time-step, every agent is given in parallel an opportunity to act. Each agent returns a set of actions, all of which are carried out in parallel, e.g. an agent can simultaneously execute a grasp, move, and turn. The size of a time-step in the world is roughly analogous to a fraction of a second. The farthest an agent can move in a single time-step is 4 units, or one third of its diameter. Typically, agents move around 1 unit per time-step.

Reinforcement Learning

Vicarious Learning

Concentric Mapping

The Simulator

The simulator for the Discotech environment was written in Common Lisp, and is run on Pentuim II machines running the Linux Operating system in the GNU Common Lisp environment. The graphical display is run in a separate process, written in Java.

Who's who in the DiscoTech

Manager on Duty: Frederick Crabbe, Professor, USNA
General Manager: Michael Dyer, Professor, UCLA
DJs: Position Open
Bartenders: Anand Panangadan and Gerald Chao, Graduate Students, UCLA

Tech Reports and Other Publications

Chao G., Panangadan, A. , and Dyer, M.G. Learning to Integrate Reactive and Planning Behaviors for Construction. in the proceedings of the Sixth International Conference on Simulation of Adaptive Behavior (SAB), 2000.
Chao, G. and Dyer M. G., Concentric Spatial Maps for Neural Network Based Navigation UCLA-TR #990016.
Crabbe, F. and Dyer, M., MAXSON: Max-Based Second-Order Neural Network Reinforcement Learner for Mobile Agents in Continuous Environments. UCLA-TR #990009.
Crabbe, F. and Dyer, M., Goal Sequence Achievement Using Higher-Order Neural Connections in Construction Agents. UCLA-TR #980018.
Crabbe, F. and Dyer, M., Cooperative Behavior, Communication and Task Decomposition in Protective Structure Construction by Neurally Controlled Agents. UCLA-TR #980019.

Acknowledgements

This work is supported in part by an Intel University Research Program grant to my advisor, Michael Dyer. The Pentium II machines were also donated by Intel. Microsoft has donated several pieces of software.