In the landscape of artificial intelligence, the definition of a "concept" has often been treated as a static mapping of features. However, in the seminal 1987 paper Formulating Concepts According to Purpose, Smadar Kedar-Cabelli argues that concepts are not merely fixed sets of attributes but are dynamic constructs shaped by the specific goals of an agent. This work represents a significant shift in symbolic AI, moving away from rigid categorization toward a functional, context-dependent understanding of the world.
The Problem: The "Chicken and Egg" of Learning
During the mid-1980s, Explanation-Based Generalization (EBG) was a dominant paradigm in machine learning. EBG allowed a system to learn a general rule from a single example by explaining why that example satisfied a specific "target concept." However, this approach suffered from a fundamental limitation: the target concept (e.g., the functional definition of a "cup") had to be provided to the system in advance.
Kedar-Cabelli identified this as a "chicken and egg" problem. If an AI must already know the definition of a cup to learn what a cup looks like, where does that initial definition come from? In reality, human concepts are fluid. A "cup" for drinking hot tea requires heat insulation, while a "cup" used as a paperweight requires only mass and stability. The paper addresses this gap by proposing a method for automatic concept formulation—deriving definitions from intended use rather than pre-coded templates.
The PurForm System
To demonstrate this theory, the author developed PurForm, a system implemented in PROLOG that formulates concept definitions based on an agent's purpose. The system operates through a three-stage process:
1. Isolating the Target Concept
The process begins with a high-level goal, such as ingested(robbie, hot_tea, X). The system identifies the unknown variable X as the target concept that must be defined to complete the plan.
2. Formulating the Purposive Functional Definition
PurForm utilizes Goal Regression, a standard AI planning algorithm, to work backward from the goal. By analyzing a plan (e.g., pouring, grasping, lifting, and drinking), the system determines the "weakest preconditions" necessary for the plan to succeed.
To prevent the resulting definition from becoming an exhaustive list of every environmental factor, Kedar-Cabelli introduced the Analyze-Role Heuristic. This heuristic filters preconditions based on two criteria:
- Relevance: The property must directly involve the object in question.
- Intrinsicness: The system distinguishes between properties the agent can change (extrinsic) and those it cannot (intrinsic). For instance, an object's location is extrinsic because the robot can move it, but its porosity is intrinsic. Only intrinsic properties are included in the concept definition.
3. Operationalizing the Definition
Finally, the system uses EBG and a single physical example (e.g., a ceramic mug) to map these abstract functional requirements—such as "must be graspable"—to observable structural features, such as "possesses a handle."
Key Insights: Concepts as Tools
The primary contribution of this research is the shift in perspective: concepts are tools. They are descriptions of objects that enable a specific plan. This framework provides a logical basis for context-sensitivity. By simply changing the agent's goal in the planner, PurForm can generate entirely different definitions for the same object class.
Furthermore, the use of Intrinsicness as a Filter provides an elegant solution to the frame problem in concept learning. By defining concepts based on the limitations of the agent’s own capabilities, the system ensures that the learned definitions are pragmatically useful rather than just descriptively accurate.
Results and Evaluation
The paper evaluates PurForm through a case study involving the formulation of a "HOT-CUP" concept. In this scenario, the system successfully derived a structural definition that was more specialized than a generic "CUP." It correctly identified that for the purpose of drinking hot tea, the object must be non-porous, made of insulating material, and light enough for the specific agent to lift.
Notably, the system demonstrated an ability to generalize across agents. By replacing a specific robot ("Robbie") with a generic "grasper-robot" in the domain theory, PurForm ensured the learned concept remained valid for any agent with similar physical constraints.
Limitations and Legacy
While pioneering, the PurForm system faces several challenges common to 1980s symbolic AI:
- Knowledge Engineering: The system relies on a meticulously hand-coded "Domain Theory." If the underlying logic rules regarding physics or the environment are incomplete, the system cannot formulate the concept.
- Artifact Bias: The purposive approach is highly effective for man-made artifacts with clear functions but is less applicable to natural categories (like "tree" or "sunset") where a singular "purpose" is harder to define.
- Single-Goal Focus: The original implementation handles objects serving one purpose at a time, leaving the integration of multi-functional objects for future research.
Despite these limitations, Formulating Concepts According to Purpose remains a landmark paper for its attempt to ground abstract concepts in the practical, goal-oriented realities of an autonomous agent.
Summary updated