A Mimetic Entity is a Machine that may mimic the personality of a human being. A Functional Response Emulation Device is an ME which synthesizes the combined behavior of agents or subsystems that share a common goal or goals into a unified representation or function, perhaps using a "human" interface with us.

I believe that "Intelligence" in the common usage is a very broad term. I would prefer to think about [Most of AI] these systems in terms of ARTIFICIAL REASONING, which is a subset of human behavior.

It seems apparent that since technology is only capable of producing isolated pieces of human behavior that the metrics may have to be tailored for each category of components. With that in mind, it seems that the first order of business would be to define these categories.

Newell listed these properties that an intelligent system must have. They are, in fact, categories of human behavior:

• recognize and make sense of a scene
• understand a sentence
• construct a correct response from the perceived situation
• form a sentence that is both comprehensible and carries a meaning of the selected response
• represent a situation internally
• be able to do tasks that require discovering relative knowledge

Yet human intelligence also has properties that are very hard to describe:

• creativity (ability to synthesize new ideas from old ones)
• attitude (determination vs. apathy for instance)
• comprehension of BOTH spoken and written language

Ray Dillinger of Neuromedia Inc. proposed some categories for future chatterbot contests in a Wired magazine article.

He says:
"I think that the Turing Test is focusing far too much attention on the process of deceiving people. I'd much prefer to see individual competitions for different events relating to the real obstacles involved -- 'best semantic comprehension of user utterances', 'most useful chatterbot', 'best ability to answer direct questions', 'best ability to infer user emotion from user utterances', 'best ability to resolve preferences to previous conversation', etc. "

I agree with his idea for these categories, but he fails to understand the nature of the deception in a Turing test.

To me, it is like this: Sensitivity Training is for people who want to *SEEM* sensitive, much like mimetic training is for robots that we want to seem human.

There is no shame in only seeming human; in fact, it is admirable to strive for this kind of advanced ergonomics. The usefulness of a program that only seems human should not be measured in terms of human intelligence, but in terms of the performance of the functional applications under it's control, and the ease with which they may be employed by us via our "human" interface with the ME. Once again, each agent or subsystem would be measured by the metrics of its particular category.

The Albus definition of intelligence is fine for systems that must use ARTIFICIAL REASONING, but our Mimetic Entities would create a unified representation to the user for these functionally nonhuman systems. The ME works in the environment that we humans work best in, i.e. working with other humans. And humans are indeed multi-faceted creatures. Likewise, in our vision for the human-computer interface of an ME, there could be no single metric to describe its behavior, since that behavior is comprised of very specialized, often goal oriented, subsystems or agents.

Regards,
Robby.

For more information on Robby Garners work please visit http://www.robitron.com/