Ron Petrick's Webpages

Workshop paper

Learning Action Effects in Partially Observable Domains, K. Mourão, R. Petrick, and M. Steedman, Proceedings of the ICAPS 2009 Workshop on Planning and Learning, pages 15-22, 2009.

[ bib | pdf ]


We investigate the problem of learning action effects in partially observable STRIPS planning domains. Our approach is based on a voted kernel perceptron learning model, where action and state information is encoded in a compact vector representation as input to the learning mechanism, and resulting state changes are produced as output. Our approach relies on deictic features that embody a notion of attention and reduce the size of the representation. We evaluate our approach on a number of partially observable planning domains, adapted from domains used in the International Planning Competition, and show that it can quickly learn the dynamics of such domains, with low average error rates. Furthermore, we show that our approach handles noisy domains, and scales independently of the number of objects in a domain, making it suitable for large planning scenarios.