Seminar on Computational Learning and Adaptation

Symbolic Function Approximation for Control

Seth Rogers
Intelligent Systems Laboratory
Daimler-Benz Research & Technology Center
1510 Page Mill Rd., Palo Alto, CA 94304

Controlling physical and simulated physical systems, such as an airplane or an automobile, generally takes the form of using a response function to compute control adjustments based on the current state of the enviroment and the desired state. These systems are frequently complex and difficult to control, requiring much specific, detailed knowledge. An approach to accurately controlling these systems is to enable the system to autonomously approximate the ideal response function through experiencing the results of its own actions. Although most work along these lines uses strictly numerical regression techniques for minimizing error, our work utilizes a symbolic high-level representation. Applying symbolic techniques to learning control of physical systems allows reasoning at a high level of abstraction and tight integration with other symbolic capabilities, such as natural language and planning. SPLICE (Symbolic Performance & Learning In Continuous-valued Environments) is a symbolic agent for function approximation and control implemented in the Soar architecture. SPLICE uses a three-level framework to first classify its sensory information into symbolic regions, then map the set of regions to a local model, then use the local model to determine an action. The agent monitors the results of the action and incrementally learns by changing its action mapping and local models. Over time, the models gradually become more specific and accurate. SPLICE performs comparably to other function approximation algorithms in a variety of domains. The final goal is to create an effective controller and improve our understanding of symbolic processing for complex physical systems.

Date: Weds., Mar. 12 Time: 4:15-5:30PM Place: Gates 100

The goal of this seminar is to increase communication among local researchers with interests in computational approaches to learning and adaptation. If you would like to be added to (or removed from) the mailing list, or if you are interested in giving a talk in the seminar, please send email to

Return to seminar schedule.