A Hill-Climbing Approach to Machine Discovery
Abstract
To account for new observations, scientists must often change an initial theory (set of beliefs) that has become inconsistent due to this data. Automated discovery systems should also be able to revise their beliefs in such scenarios, in order to regain a consistent database. To this end we have constructed REVOLVER, a program that performs both discovery and belief revision. When the system makes erroneous inferences, it employs hill climbing to search for a new consistent set of beliefs. In this paper, we first discuss the goals and tasks addressed by REVOLVER. Next we describe the program's operation, using an example from chemistry to illustrate the system's representation, its basic rules and inference process, and its method for belief revision. We then evaluate the system, showing its generality (through its replication of discoveries in natural domains) and its robustness (through its ability to run efficiently in artificial domains). Finally, we discuss related work on belief revision and ideas for future work.
Cite
Text
Rose and Langley. "A Hill-Climbing Approach to Machine Discovery." International Conference on Machine Learning, 1988. doi:10.1016/B978-0-934613-64-4.50042-6Markdown
[Rose and Langley. "A Hill-Climbing Approach to Machine Discovery." International Conference on Machine Learning, 1988.](https://mlanthology.org/icml/1988/rose1988icml-hill/) doi:10.1016/B978-0-934613-64-4.50042-6BibTeX
@inproceedings{rose1988icml-hill,
title = {{A Hill-Climbing Approach to Machine Discovery}},
author = {Rose, Donald and Langley, Pat},
booktitle = {International Conference on Machine Learning},
year = {1988},
pages = {367-373},
doi = {10.1016/B978-0-934613-64-4.50042-6},
url = {https://mlanthology.org/icml/1988/rose1988icml-hill/}
}