next up previous
Next: Example Up: An Open Agent Architecture Previous: Comparison with Other Agent


In our earlier scenario, the mail agent was rather limited. To test our user interface and agent architecture more fully, we are creating a more substantive mail management agent, MAILTALK.

  It has become common to develop mail managers that manipulate messages as they arrive according to a set of user-specified rules. The virtue of such systems is that users can make mail management decisions once, rather than consider each message in turn. However, a number of problems exist for such systems, as well as for all agent systems that we know of, especially when considered as tools for the general population.

Our prototype MAILTALK was built to address these concerns.

Rule specification.
Based on technology developed for the SHOPTALK factory simulation system [[2], [3], [4]], MAILTALK permits users to specify rules by describing complex invocation conditions, and arguments with a multimodal interface featuring typed and spoken natural language, combined with direct manipulation. For example, the user can delegate to the mail agent as follows: ``When Jones replies to my message about `acl tutorials', send his reply to the members of my group." Here, Jones's reply cannot be selected or pointed at since it does not yet exist. The English parser produces expressions in the temporal logic, which are evaluated against various databases (e.g., the mail database, or a simulation database).

Predicting behavior.
By giving end users the power to write their own rules means we have given them the freedom to make their own mistakes. Before letting a potentially erroneous collection of agents loose on one's mail (or, more generally, the network), we encourage users to simulate the behavior of those agents. Included with MAILTALK is a knowledge-based simulation environment that allows users to create hypothetical worlds, and permits them to send test messages or re-examine old mail files. In response, the system fires the relevant rules, and updates a simulation database with the events that have happened. This database can extend the actual mail file, permitting expressions that depend on the entire database to be evaluated (e.g., ``when more than 5 messages from cohen are in < point to icon for mail file>, move them to <icon for `unimportant mail'>).

Reviewing History.
In order to determine if the resulting behavior was in fact desired, users can ask questions about the results of the simulation, can view the simulation graphically, and can rewind the history to interesting times (e.g., when a message was read, or when a message was forwarded to a member of a given mail group). When satisfied with the resulting behavior of the collection of rules, users can install them in the real world to monitor the real mail file. Moreover, users can ask questions about the real mail database, such as ``Who has replied to my message of November 26 about budgets?''

next up previous
Next: Example Up: An Open Agent Architecture Previous: Comparison with Other Agent

Adam Cheyer
Mon Aug 12 15:12:15 PDT 1996