My PhD dissertation is about development and the structure of activity. How could a neonate, with very little innate behavioral knowledge, experiment with and learn about its motor capabilities and over time organize them into activities that can be used to achieve goals?

Central to my hypothesis of a plausible developmental account:

• the ability to model (learn) how sensors behave in the context of primitive action
• the ability to examine traces of past behavior and reason about why each action in the trace unfolded as it did
• the ability to use this reasoning to extract hypothesis plans from traces of past behavior and attempt to confirm that they work
• a high level control system that attempts first to satisfy basic needs, and once those are satisfied to pursue learning goals
• a set of primitive reflexes that ensure the neonate gains experience with important action outcomes

In my dissertation, I motivate the approach and explain in detail an implementation of a system that develops activities in a simple testbed environment.

download the dissertation (1.29M, pdf)

• Matthew D. Schmill, and Paul R. Cohen. "A motivational system that drives the development of activity". In Proceedings of the First International Joint Conference on Autonomous Agents and Multi-Agent Systems, Bologna, Italy, July 2002.
• Oates, Tim, Matthew D. Schmill, Paul R. Cohen. "A Method for Clustering the Experiences of a Mobile Robot that Accords with Human Judgements". In Proceedings of the Seventeenth Interational Conference on Artificial Intelligence, pp. 846-851. AAAI Press/The MIT Press: Menlo Park/Cambridge, 2000.
• Schmill, Matthew D. Schmill, Tim Oates and Paul R. Cohen. "Learning Planning Operators in Real-World, Partially Observable Environments". In Proceedings of the Fifth International Conference on Artificial Intelligence Planning and Scheduling pp. 134-135. ACM: New York, NY, 2000.