4.8 How Agents Reason

Men and machines such as agents reason through simple to elaborate networks of rules:

      IF     X,     AND     Y,     THEN        Z

Some of these rules are codified from the domain of experts; hence, the development of expert system in the early 1990s. However, these systems fell out of popularity after some initial enthusiasm when they proved to be expensive to maintain and brittle in deployment. Expert systems represented a set of rules in such areas as making soup or configuring systems or auditing tax returns. Some expert systems still exist; for example, the TriPath medical system uses rules that it developed from pathologists to examine Pap smears to diagnose for cervical cancer in its FocalPoint system.

The FBI and IRS both set up AI labs with the intent of developing expert systems to assist field agents with working cases and developing good prosecutions. The idea was to codify the experience of seasoned FBI agents who had worked and solved specific types of criminal cases. These expert systems would subsequently aid younger agents in working cases to prosecution and eventual conviction.

The IRS had various applications under development, most of which never left the lab, due in part to the high maintenance cost of the expert systems. For example, one application was to automate the audit examination process; unfortunately, the tax code and forms change every year, with Congress cranking out new legislation, meaning the rules of the expert system would need to be constantly changing. Audits can involve multiple years, meaning the expert system would have to incorporate hundreds of rules from each of those years. In the end, the task proved to be expensive to maintain.

However, there is a different method by which rules can be constructed; this involves data mining. Replacing expert systems as reasoning engines was the development of neural networks and machine-learning algorithms in the area of AI. Rather than developing rules from experts and taking a top-down approach to knowledge acquisition, rules can be extracted from observations in large databases. This is the inductive method of data analysis, now known as data mining, which uses machine learning and is a bottom-up approach to knowledge acquisition.

These processes of rule creation are not mutually exclusive; in fact, a hybrid system is probably the ideal solution for investigative data mining applications, in which some rules are drawn from years of investigators' experience, coupled with rules extracted from hundreds of thousands of cases from large databases. This type of man-machine hybrid system is the topic of a proposed data mining architecture in Chapter 11. Agents, as engines of inference, can use both types of rules. To develop intelligence in agents, certain steps can be taken. Briefly, they involve the following type of rule sequencing and construction:

The user or developer provides a set of rules that describe a desired behavior: When X happens, then do Y. This can be done using a plain-text editor and then transcribed to code-such as C or Java.
The reasoning system is next provided with a set of conditional input events, such as When a match of Entity Z898R from List DEA-01/02/04 happens, do Y.
The reasoning system is provided with interfaces to perform or initiate various desired actions; for example, do Y may require that an alert be made by sending a message to a system object, by writing a file, or by other system action that a program can perform.
After the reasoning system is initiated, it can wait for an event to arrive. It will extract facts from the event and then evaluate its rules to see if the new facts cause any of them to fire. If one or more rules fire, it may cause additional action to be initiated or a record to be written or updated.

The above process follows a set structure, leading to the creation and use of conditional rules and logic, which can be coded in a variety of ways. Here is an example:

      IF      (Condition 1)      OR      (Content A)      AND     (Condition 3)      THEN (Action Z)

This can be demonstrated by the example of a system for issuing alerts to, say, customs agents at point-of-entry stations, based on conditions gleaned from a plate number input into a network system using models developed from both human investigators' experience and machine-learning-generated rules. These data mining rules could well have been developed from an extensive analysis of prior convicted cases of contraband prosecutions:

      Condition fields:        IF INSURER is None (Condition 1)           Source: Human Domain        OR YEAR is 1988 (Content A)           Source: DMV Registration Record        AND MAKE is CADILLAC (Condition 3)           Source: Data Mining Model       Prediction # 1: THEN ALERT is Medium (Action Z)           Inspect Trunk