Symptom Management for Schizophrenic Agents

AAAI 1996

TL;DR

This paper tackles the challenge of building complex autonomous agents by proposing a "socially situated AI" approach. Instead of just focusing on an agent's internal behaviors, it suggests managing the effect of those behaviors on the user, thereby creating smoother, more understandable agent interactions and overcoming the "schizophrenia" of abrupt behavioral shifts.

What Problem Does This Paper Address?

Building sophisticated autonomous agents is difficult because as more behaviors are combined, their interactions become incredibly complex and hard to manage. This often leads to a phenomenon the author calls "schizophrenia" in agents: they exhibit stereotyped behaviors with abrupt, jarring switches between different modes. This breakdown is attributed to an "atomization" methodology, where complex phenomena are reduced to simple, independently interacting parts. While this atomization is necessary for practical agent design (making them easier to build and debug), it creates the problem of fragmented behavior. The paper argues that current AI approaches often ignore a crucial part of an agent's environment: its social context, including the designer and the user. This oversight obscures the root of technical problems and makes them harder to solve.

Methods

The paper proposes a new AI methodology centered on "socially situated AI." The core idea is to shift the focus from managing an agent's internal behaviors to managing the effect of those behaviors on the user. This involves creating an architecture where "atoms" are not just behaviors but also "behavior transitions." These transitions are special behaviors that smoothly change the agent from one high-level activity to another. By implementing these switches as behaviors themselves, the agent's overall interaction appears more natural and less abrupt to the user. The author is developing solutions to potential technical challenges like code explosion and complex interdependencies by integrating the agent's social context and its impact on the user into the design process.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • "Schizophrenia" as a symptom of atomization: The paper reframes the abrupt switching of agent behaviors not as a bug, but as a predictable outcome of breaking down complex systems into too-simple, independently interacting parts.
  • The importance of the social environment: The author argues that ignoring the social context (designer, user) in agent design is a critical oversight that hinders problem-solving.
  • Managing user perception over internal states: The key innovation is shifting the design goal from controlling an agent's internal "state" to managing how the user perceives the agent's actions.
  • Behavior transitions as smooth bridges: Introducing "behavior transitions" as a new architectural component allows for seamless shifts between different agent activities, masking the underlying atomization.
  • Empowering designers with effect management: The proposed methodology aims to give designers control over the agent's effect on the user, rather than just its behavior, leading to more coordinated and understandable agents.

Key Results

The paper is primarily a proposal for a new methodology and architecture. It outlines the potential benefits of this approach, such as enabling designers to leverage the advantages of atomization (modularity, clean code, understandability) without suffering from the drawback of behavioral schizophrenia. The author aims to demonstrate that this behavior-transition architecture can lead to larger, more coordinated agents. Specific quantitative results are not presented as this is a conceptual paper.

Limitations and Future Work

The author acknowledges potential technical problems with behavior transitions, including a quadratic explosion of code, large amounts of state to manage between behaviors, and the risk of creating complicated, inter-related, and undebugable code. The future work involves developing solutions to these problems based on the socially situated approach, focusing on managing the agent's effect on the user. The author hopes to show that this architecture will allow for the creation of larger, more coordinated agents by overcoming the limitations of current atomistic approaches.

Summary updated

BibTeX
@inproceedings{1996symptom,
  title = {{Symptom Management for Schizophrenic Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure