Abstract
In the late 1970s, the development of expert systems was hindered by the "knowledge acquisition bottleneck"—the labor-intensive process of manually encoding human expertise into machine-readable code. The Instructable Production System (IPS) Project was an ambitious research initiative designed to overcome this hurdle by creating AI agents capable of learning through direct human instruction. This paper reviews the various "kernel" architectures explored during the project, identifying the essential functional components required for a machine to autonomously grow its knowledge base through interaction, observation, and self-reflection.
Introduction
The central challenge of the IPS Project was to move away from static programming toward a model of "machine growth." The goal was to build a system that could accept instruction in restricted natural language, observe its environment, and modify its own procedural rules based on feedback. By treating the computer as a student rather than a passive tool, researchers sought to create agents that could accumulate knowledge over a lifetime, becoming increasingly efficient through practice.
The Production System Foundation
The project utilized Production Systems (PSs) as its foundational architecture. In a PS, knowledge is represented as an unstructured collection of "production rules" (condition-action pairs). The system operates on a "recognize-act" cycle, where the current state of Working Memory (WM) is matched against Production Memory. When a rule's conditions are met, its action is executed, potentially altering the WM or the external environment. This modularity made PSs an ideal testbed for incremental learning, as new rules could be added or modified without rewriting the entire system.
Architectural Variations: The Kernels
Throughout the project, several distinct "Kernel" approaches were developed to test different theories of instruction and knowledge organization:
- Kernel Version 1: Implemented a "means-ends" strategy. Rules were categorized into functional types, such as means (how to achieve a goal), success recognizers, and failure recognizers.
- Additive Successive Approximations (ASA): This approach focused on "pointing" at environmental items to build fragments of methods. These fragments served as a scaffold that could be refined through subsequent interactions.
- Analogy (ANA): Designed to solve novel problems by mapping them onto existing, hand-coded solutions. It utilized a semantic network to find paths between new problem states and known domains.
- Problem Spaces: This version modeled all agent behavior as a search through a space of states and operators, providing a formal structure for goal-oriented learning.
- Schemas: A slot-filling approach was used to manage complex tasks. Different slots handled initialization, execution, and monitoring, allowing the system to group related rules into cohesive units.
- Kernel Version 2 & IPMSL: These focused on dynamic, step-by-step interaction. The instructor would "step" the system through a task, and the system would build new rules by observing the instructor's choices and the resulting state changes.
The Eight Dimensions of Instructability
A primary contribution of the IPS Project was the identification of eight functional dimensions that define the design space for an instructable agent:
- Interaction: The ability to communicate with a human teacher.
- Organization: The structure used to group and manage individual rules.
- Explanation: The system's ability to reveal its internal reasoning to the teacher.
- Accommodation: The process of incorporating new rules into the existing memory.
- Connection: Linking new knowledge to existing concepts and goals.
- Reformulation: The ability to translate high-level instructions into executable rules.
- Evaluation: Assessing the effectiveness of learned rules.
- Compilation: Optimizing slow, interpretive knowledge into fast, automatic procedures.
Key Insights and Mechanisms
The Centrality of Explanation
For effective instruction to occur, the machine must be transparent. If an agent makes an error, it must be able to explain why it chose a specific action. This allows the instructor to identify the faulty rule or the missing piece of information. Without explanation, the instructor is forced to guess the agent's internal state, making the teaching process inefficient.
Active vs. Passive Learning
The project highlighted a shift from passive systems—which wait for input—to active agents. Active agents evaluate their own progress, identify gaps in their knowledge, and ask the instructor specific, targeted questions. This reduces the burden on the teacher and ensures that instruction is relevant to the agent's current needs.
The Compilation Step
Knowledge acquired through instruction is often "interpretive" and computationally expensive to execute. The IPS researchers identified the need for a "compilation" phase. Through practice, the system should be able to collapse complex sequences of rules into single, optimized productions, achieving the kind of automaticity seen in human skill acquisition.
Limitations and Conclusion
While the IPS Project successfully implemented several kernels (Kernel 1, ANA, Kernel 2, and IPMSL), it did not produce a single, perfectly integrated agent. Major limitations included the high "instructor burden," where the human teacher still had to perform significant cognitive work to format information for the system. Additionally, the restricted natural language interfaces remained a bottleneck for fluid communication.
Despite these challenges, the project provided a vital conceptual map for the future of machine learning. It demonstrated that the Production System architecture is uniquely suited for exploring the multifaceted nature of knowledge acquisition, laying the groundwork for modern research into autonomous agents and interactive AI.
Summary updated