A Model of Creative Understanding

Kenneth Moorman, Ashwin RamAAAI 1994

Introduction

In the field of Artificial Intelligence, creativity is frequently conceptualized as a generative tool—a means to produce novel solutions, art, or text. However, this paper posits that creativity is equally fundamental to the process of understanding. While "mundane" comprehension involves mapping new information onto existing knowledge structures, "creative understanding" is required when an agent encounters concepts that fundamentally defy its internal database. This is particularly evident in science fiction, where narratives often introduce revolutionary novelty (R-Novelty) that requires the reader to construct entirely new conceptual models on the fly.

The ISAAC System

To explore this, the authors developed ISAAC (Integrated Story Analysis And Creativity), a system designed to comprehend narratives that introduce alien concepts. ISAAC moves beyond simple pattern matching by employing two primary mechanisms: the Cycle of Creative Understanding (CUP) and Function-driven Morphological Synthesis (FMS).

The Cycle of Creative Understanding (CUP)

When ISAAC encounters a concept it cannot immediately resolve through standard memory retrieval, it initiates the CUP algorithm. This is a hierarchical fallback process that attempts increasingly radical methods to achieve comprehension.

The Cycle of Creative Understanding (CUP) fallback algorithm, illustrating the sequence of retrieval, incorporation, analogy, and reformulation.

The cycle begins with standard Memory Retrieval. If this fails, the system attempts Incorporation, looking for relationships between known concepts and the new input. If no direct base exists, it tries Baseless Analogy. Finally, if all else fails, it resorts to Problem Reformulation via FMS. In this final stage, the system actively merges and manipulates attributes of known concepts to invent a new explanatory model that fits the narrative context.

Function-driven Morphological Synthesis (FMS)

FMS is the "creative engine" of the system. It uses the "function" of an object as an invariant anchor. For instance, to understand a sentient robot in a story, ISAAC identifies the desired function (a "willful agent") and looks for existing concepts that fulfill it (such as a "human"). It then grafts the functional attributes of a human onto the physical attributes of a robot, creating a novel "story-robot" concept.

Heuristics and the "Bizarre" Threshold

A significant challenge in creative AI is preventing the system from making illogical or "too bizarre" leaps. To manage this, ISAAC utilizes a Multidimensional Knowledge Grid.

The Multidimensional Knowledge Grid, categorizing knowledge across physical, mental, social, emotional, and temporal domains.

Knowledge is organized into domains (Physical, Mental, Social, etc.) and categories (Agents, Actions, Objects, States). Moving between cells in this grid incurs a computational "cost." This heuristic guides the system toward interpretations that are creative yet plausible. For example, treating a machine as a social agent involves a "grid shift" that the system tracks. If the total cost of these shifts exceeds a certain threshold, the interpretation is deemed too bizarre and is rejected.

Case Study: "Men Are Different"

The system's efficacy was tested using Alan Bloch’s short story, "Men Are Different." The story is narrated by a robot archaeologist who accidentally kills a human because he assumes the human is a robot with a "shot thermostat."

Initially, ISAAC failed to understand the narrator because its internal definition of a robot was a non-sentient industrial tool. By applying FMS, ISAAC successfully merged the concept of a "man" (a volitional agent) with an "industrial robot" (metal frame, sensors). This allowed the system to not only follow the plot but also recognize the story's central irony: the tragedy arises from the robot narrator mistakenly applying its own mechanical "repair" logic to a biological entity.

Conclusion and Future Work

The ISAAC system demonstrates that creativity is a vital component of deep comprehension. By shifting the focus from generation to understanding, the authors provide a framework for AI that can appreciate complex literature and novel ideas. Future work will focus on refining evaluation metrics for creative performance and exploring how these temporary "story concepts" might permanently evolve an AI's long-term knowledge base.

Summary updated

BibTeX
@inproceedings{moorman1994model,
  author = {Kenneth Moorman and Ashwin Ram},
  title = {{A Model of Creative Understanding}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}

Loading related content...

Figure