CASAP: A Testbed for Program Flexibility

TL;DR

This paper introduces CASAP, a system designed to create more flexible knowledge-based systems for performing actions. By combining procedural representations of actions with a dynamic information retrieval subsystem, CASAP aims to simplify the specification of new actions and enhance their ability to acquire necessary information from diverse environments.

What Problem Does This Paper Address?

The paper addresses the limitations of early intelligent systems, such as theorem provers and the General Problem Solver (GPS), which struggled with general problem-solving due to inadequate mechanisms for ordering operators and incorporating domain-specific knowledge. While later systems like PLANNER improved this by using pattern-directed invocation and procedural representations, they often sacrificed "awareness," making it difficult to reason deductively about actions or their consequences. The authors highlight two key problems with the procedural approach: the rigid transfer of control and data between routines, and the difficulty in coordinating multiple procedures to act in concert.

Methods

CASAP attempts to bridge the gap between procedural knowledge representation and general problem-solving by introducing a novel system organization. Instead of direct subroutine calls, CASAP uses a single information retrieval routine that acts as an intermediary between any action-performing routine and the global data base. This means routines only need to know what information they require, not where or how to find it. Furthermore, CASAP utilizes a "context" mechanism, established and maintained by the information retriever, which influences how information is retrieved. This context can dynamically modify the behavior of invoked routines by shaping the information they receive. The system also uses English for specifying actions and concepts, though input is pre-parsed.

Datasets

Not discussed in the paper. The paper focuses on the system's architecture and a demonstration using a card game (Hearts) as an example.

Key Insights and Ideas

  • Decoupling Information Needs from Retrieval: The core idea is that routines should declare their information needs, and a dedicated retriever handles the complex task of finding that information, abstracting away the data base's structure.
  • Context-Driven Information Retrieval: The information retriever dynamically adjusts its search based on a "context," allowing for more flexible and situation-dependent information acquisition. This means the same request can yield different results depending on the current state of the system.
  • Procedural Knowledge with Enhanced Awareness: CASAP aims to retain the power of procedural representations while mitigating the loss of deductive reasoning capabilities seen in earlier systems.
  • Flexibility through Dynamic Information Gathering: The system's ability to dynamically obtain information as needed is presented as a key factor in achieving flexibility, mirroring human-like information utilization.

Key Results

The paper demonstrates CASAP's flexibility through a detailed example of playing the card game Hearts. The system, starting with no knowledge of cards or games, successfully interprets commands like "Insert card X in the hand of Player Y" and "Remove object hearts which highest, from hand of Player1." The interpretation of these commands, especially the "pass the highest heart" action, showcases how the system dynamically retrieves necessary information (like card suits and values) and uses context to resolve ambiguities and perform actions. The authors emphasize that the intelligence lies not just in the routines themselves, but in the dynamic interplay between routines, the information retriever, and the context.

Limitations and Future Work

The authors acknowledge that the context mechanism in CASAP is currently limited. It primarily influences information retrieval and does not directly affect action execution when information is not explicitly required. Furthermore, it's an "information-based" rather than an "action-based" influence. The paper suggests that moving towards languages that are "context-dependent for the interpretation of both actions and data" is a crucial next step in programming language advancement. The "modification component" for altering actions based on context is also noted as currently null, indicating a potential area for future development.

Summary updated

BibTeX
@inproceedings{1973casap,
  title = {{CASAP: A Testbed for Program Flexibility}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1973},
}
Figure