| It is generally accepted that important advances in Artificial Intelligence will require significant amounts of knowledge. Machine learning techniques were developed to acquire large amounts of knowledge. However, it was reported in 1988, that the costs involved in using the increasing amounts of knowledge eventually consume the benefits of the additional knowledge making further learning detrimental to performance. Almost all approaches to the “Utility Problem” have been to restrict or slow learning. This research investigates and characterizes long-term learning using Soar. Learning was discovered to eventually either stop when the domain was finite and had been effectively “solved” or settle to a constant rate when the domain was large or infinite. In addition, many of Soar's chunks were found to be never used or to become unused. Removal of chunks based on demonstrated low use is found to improve performance in a complex domain over the long term. The Utility Problem is shown not to be a natural limit but an artifact of monotonic learning in large or infinite problem domains. |
