In many learning and optimization tasks, the sample cost of performing the task is prohibitively expensive or time consuming. For this reason, attempting to directly employ a learning algorithm on the task could quickly become intractable. For this reason, learning is instead often performed on a less expensive task that is believed to be a reasonable approximation or a surrogate of the actual target task. This paper serves to present and motivate the challenging open problem of simultaneously performing learning on an approximation of the true target task, while at the same time adapting the surrogate task used for learning to be a better representation of the true target task. Our work, which is still in progress, is performed in the RoboCup 3D simulation environment where we attempt to learn the configuration parameters for an omnidirectional walk engine used by humanoid robot soccer playing agents.

PDF, PS, HTML

In GECCO Surrogate-Assisted Evolutionary Optimisation (SAEOpt) Workshop, Denver, Colorado, USA, July 2016.

@inproceedings{SAEO16-MacAlpine, title={Adaptation of Surrogate Tasks for Bipedal Walk Optimization}, author={Patrick MacAlpine and Elad Liebman and Peter Stone}, booktitle={GECCO Surrogate-Assisted Evolutionary Optimisation (SAEOpt) Workshop}, month={July}, address={Denver, Colorado, USA}, url="http://www.cs.utexas.edu/users/ai-lab?macalpine:saeo16", year={2016} }

Slides (PDF)