The essential dynamics algorithm combines ideas from the adaptive control and policy search. The resulting algorithm works in high dimensional, continuous domains. The papers below describe the algorithm, relate it to Linear Quadratic Control and Reinforcement Learning, and describe it on two tasks: a toy bicycle riding problem, and controlling a robot arm on a Segway base, which dynamically rotates.
- The Essential Dynamics Algorithm: Fast Policy Search in Continuous Worlds, Martin, M.C. Media Laboratory Vision and Modeling Technical Report 582, MIT. June 2004.
- Controlling Cardea: Fast Policy Search in High Dimensional Space, Martin, M.C. Media Laboratory Vision and Modeling Technical Report 583, MIT. June 2004.
- The Essential Dynamics Algorithm: Essential Results, Martin, M.C. Artificial Intelligence Laboratory Memo, MIT. May 2003.
The source code for the bicycle riding experiments.