DAYDREAMER

DAYDREAMER is a cognitive architecture developed at UCLA by Erik Mueller. It models the human stream of thought and its triggering and direction by emotions, as in human daydreaming. The architecture is implemented as 12,000 lines of Lisp code.

Summary of DAYDREAMER

Example daydream

A sample daydream generated by DAYDREAMER is as follows:

REVENGE1

I study to be an actor. I am a movie star even more famous than
he is. I feel pleased. He is interested in me. He breaks up with
his girlfriend. He wants to be going out with me. He calls me up.
I turn him down. I get even with him. I feel pleased.

Concerns and emotions

In response to external and internal circumstances, DAYDREAMER activates, processes, and terminates multiple concerns. Each concern is an instance of a personal goal or ongoing life objective of the program.

Concerns are motivated by emotions, which determine what concern to process at any given time. The strengths of these emotions, although initially set according to the intrinsic importances of particular personal goals, are subject to dynamic modification as unexpected consequences of a concern are recognized.

Generation of daydream scenarios

The sequences of both fanciful and realistic events that make up daydream scenarios are produced using the building blocks of planning rules, inference rules, and episodes. Planning rules specify methods of varying degrees of plausibility for breaking down a subgoal into further subgoals, whereas inference rules specify consequences of various situations.

Daydream scenarios are generated by the planning mechanism, which repeatedly applies planning and inference rules to a selected concern. The planning mechanism is employed to generate possible behaviors of the daydreamer as well as the possible behaviors of others.

Mutation

A mechanism for modifying existing daydream scenarios is the mutation of the objectives of unsuccessful action subgoals.

Episodes

Episodes are aggregates of rule instances applied in a concrete real or imagined situation. Some episodes are hand-coded and provided to the program as input, whereas others are generated as the program daydreams and interacts in the simulated real world. Episodes are indexed in episodic memory under subgoal objectives, emotions, persons, and other features. Once retrieved, past episodes may be applied, in whole or in part, to a new concern by the analogical planning mechanism. Episodes reduce search in planning and enable scenario details to be filled in.

Serendipity

The serendipity mechanism recognizes the unexpected applicability of some possibility related to one concern, to another active concern. The serendipity mechanism conducts an intersection search, from a point associated with a new possibility to a point associated with an active concern, through the space of currently accessible planning rules. Found paths are then verified by progressive unification and employed through analogical planning. Once a serendipity occurs, the resulting plan is stored in episodic memory so that in the future a similar plan can be generated without having to chance upon a similar serendipity.

Daydreaming goals

A collection of daydreaming goals augment the program's personal goals and initiate useful daydreaming activity. The daydreaming goals of rationalization (generating scenarios to rationalize a failure), roving (shifting attention from an unpleasant failure), and revenge (generating scenarios in which revenge is attained) model the daydreaming that humans perform in order to reduce negative emotional states. The daydreaming goals of reversal (generating scenarios in which a past or imagined future failure is avoided), recovery (generating scenarios in which a goal that failed in the past succeeds in the future), rehearsal (generating possible scenarios for achieving an active goal), and repercussions (exploring and planning for hypothetical future situations) enable the program to improve its future external behavior, that is, to learn.

Learning through daydreaming

Learning through daydreaming is accomplished by adding daydreamed episodes to memory: Various alternative scenarios involving a given situation are generated, evaluated as to their realism and desirability, and stored as episodes. When a similar situation arises in the future, the best and most similar retrieved episode is applied to that situation through analogical planning. Generation of hypothetical future scenarios improves the behavior of the program because negative consequences of various courses of action can be detected in advance and thus avoided.

Learning is also accomplished by adding new planning and inference rules: In response to a side effect real or imagined failure, the reversal daydreaming goal determines what actions might have been taken in order to avoid that failure. Rules are then created to anticipate similar failures in the future and carry out appropriate actions to prevent those failures.

History

DAYDREAMER was begun by Erik Mueller in 1983 and completed in 1987. It was followed by the ThoughtTreasure program, which was started in 1993.

External links

DAYDREAMER code

References

Mueller, Erik T. (1990). Daydreaming in humans and machines. Norwood, NJ: Ablex.
Mueller, Erik T., & Dyer, Michael G. (1985). Towards a computational theory of human daydreaming. Proceedings of the Seventh Annual Conference of the Cognitive Science Society. Hillsdale, NJ: Lawrence Erlbaum.
Mueller, Erik T., & Dyer, Michael G. (1985). Daydreaming in humans and computers. Proceedings of the Ninth International Joint Conference on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann.