Learning How to Do Things: Papers from the AAAI Fall Symposium
Mathias Bauer and Charles Rich, Cochairs
Knowing how to do things is an important category of knowledge underlying many kinds of intelligent behavior in artificial agents, such as critiquing, advice giving, tutoring, collaboration, and delegation. In the current state of the art, most of this procedural knowledge is encoded manually by a single person (or a small team) who needs to be expert in both the task domain and the appropriate knowledge representation formalisms. This is a serious bottleneck in the development of these kinds of systems. The focus of this symposium is on how to automate or partially automate the acquisition of procedural knowledge, namely, indexed collections of what are variously called macros, plans, procedures, or recipes for action. The techniques for acquiring this knowledge may depend on many variables, including: size of the domain (e.g., number of recipes); amount of input data; number of steps in a typical task; type of tasks (e.g., analysis versus syn-thesis); number of agents involved (e.g., one, two, or many); type of agents involved (e.g., human versus computer); intended use of the knowledge (e.g., acting, critiquing, etc.); degree of supervision (e.g., teaching versus unsupervised learning); level of abstraction (e.g., primitive operations versus high-level goals); degree of initiative (e.g., learning by experimentation versus passively).