Approaches to Authoring of Rules for Intelligent Agents

The problem we address is practical design of methods for users, especially relatively non-technical end users, to author rules that drive rule-based intelligent agents. We are concerned especially with networked applications of intelligent agents in the realm of information access, e.g., news, mail, Web pages, and the like. For these applications, an important value of intelligent agents is their ability to customize or personalize the behavior of an application. A key challenge is how to enable an end-user to specify the behavior that is desired of that application agent. In other words, how shall the user instruct the agent? Generally, one can imagine a spectrum of approaches, ranging in complexity from full-blown programming languages, such as C++, at one end of the spectrum, to menus and direct manipulation at the other end. Scripting and macro languages, which recently have been receiving a great deal of attention and excitement in both industry and academe, occupy a place in this spectrum intermediate between full-blown programming languages and menus. This spectrum can be viewed as a trade-off frontier: lower skill and effort required vs. greater complexity and power of specified procedures/processing. Rules and inferential reasoning occupy another place in this spectrum. We place them intermediate between scripting and menus: lower skill and effort required than scripting, but more than menus; greater complexity and power of processing than menus, but less than scripting. A dream of many (we share it) is to warp out the above trade-off frontier of explicit instruction: by relying on machine learning to perform powerful implicit instruction. Pragmatically, however, we believe it is important as an incremental first step to facilitate explicit instruction, for several reasons. First, explicit instruction is important as a target for learning by induction or advice. Experience in the field of knowledge-based systems teaches us that it is usually important to first build the performance element and understand how it would exploit knowledge supplied by a learning element, before actually hooking up a learning element. Second, explicit instruction is important to augment learning: to edit the results of learning, and to complement it when only some areas of knowledge can be learned. Third, explicit instruction is important to trust learning: e.g., to confirm or inspect suggestions made by a learning element. Within explicit instruction, we believe that yes/no rules are an important first step. By yes/no rules, we mean rules that have a truth value of either true or false, as opposed to probabilistic-flavored rules (we would include fuzzy rules as in this category) that may have intermediate degrees of truth, e.g., 0.7. Yes/no rules have two advantages over probabilisticflavored rules. First, their fundamentals are well understood, notably in terms of practical experience with knowledge-based systems over especially the last fifteen years. Second, yes/no rules enable the control of tasks with predictability. (Software today is mostly written using non-randomized control constructs.)