Problems in Building an Instructable Production System
Proceedings of the Fifth International Joint Conference on Artificial Intelligence. Knowledge Acquisition. Pages 337--1.
© 1977 International Joint Conferences on Artificial Intelligence
Abstract
- Forgy, C. and McDermott, J., 1977. "OPS : A domain-independent production system language", in Proc. IJCAI-77. IJCAI
- Newell, A., 1977. "Notes on knowledge representation aspects of production systems", in Proc. IJC!ll-77.
- Moore, J. and Newell, A., 1973. "How can MERLIN understand?", in Gregg, L., Ed., Knowl edge and Cognition, Potomac, MD: Lawrence Erlbaum Associates, Pp. 201 -252.
- Rychener, M.' D. and Newell, • A., 1977. "An instruct able production system : Basic design Issues", in D. A. Waterman and F. Hayes-Roth, Eds., Po.ttern-Directeci Inference Systems, New York, . NY: Academic Press. Forthcoming. Knowled~e Acq.-2: Rychener
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
PROBLEMS IN BUILDING AN INSTRUCT ABLE
PRODUCTION SYSTEM
M. Rychener, C. Forgy, P. Langley, J. McDermott, A. Newell, and K. Ramakrishna
Department of Computer Science Carnegie-Mellon University Pittsburgh, PA 15213
Keywords: production syst.em, instruction, means-ends analysis, general intelligence, representation.
The Instructable Production System project is explor ing the incremental growth properties of production systems (PSs) by constructing a generally intelligent pr oblem-solving system by gradual (external) instruction. The definition of PS and our current architecture are given elsewhere in this volume (Newell, 1977; Forgy and McDermott, 1977). The present task domain is an abstract job shop, in which finished goods are made from raw materials. We start with a Kernel (a small PS of about 200 productions} which has ·the basic capabil ities to grow by instruction: (1} process . a restricted natural language; (2} form productions from its input; (3) impose PS control conventions on them; and (4). perform basic manipulations in its env ironment (Rychener & Newell, 1977). We lake the basic computational and representational adequacy of PSs for AI programs as established.
This short note presents some immediate difficulties we expect to encounter. These derive from the instructional situation: (1) The instructor can observe the system in the environment and can communicate with it f reel y , but cannot examine its internal structure directly. (2 } Interaction with the system is in an external language, analogous to natural language. (3) The initiative for inter action is mixed. (4) Instruction may be on any topic: specific tasks, general properties of tasks, the language of communication, possible errors, how to plan and explore, etc. (5) Knowledge and system structure gained through instruction accumulates over 'the life of the system.
Our cur rent approach uses means-ends analysis as the basic philosophy of both problem-solving and instruction. Goals are symbol structures in Working Memory that describe desired states and processing entirely through the means and tests. Means are encoded as productions that recognize goals and assert subgoals whose satisfaction will achieve the goals. Tests are encoded as productions that recognize the conditions of sati sfaction of the goal. The means productions form a means-ends network of goals. Instruction consists of elaboratin.g the nodes. of this network as required by .a t ask.
Now for the difficulties on the immediate horizon:
- Contact : How can initial contact be made with exi sting knowledge that might be relevant to the task at hand, which is not part of the means-ends network deliberately created by instruction for the task? Use of the data acquired through experience is required in any intelligent agent. Once detected, much processing can be spent on discovering relevance, but initial contact may be extremely difficult. Any general intelligent system will have too much knowledge to consider exhaustively. PS architectures exploit this by storing all knowledge as productions which are evoked only if their conditions · "see
- themselves" in the Working Memory. But means productions are acquired in specific contexts and their conditions become keyed to specific goals and task features. One approach might be to generate variations o~ current Working Memory goals and data until something is evoked.
- Incoherency: The PS may . be essentially incoherent in describing its situat ion and difficult ies to an external instructor. The means-ends network helps (by provi ding the same level of explanatory capabi.lity as in current expert systems}, but is not sufficient. E.g., describe what went wrong from the debr is left in Working Memory. The PS's diagnostic and explanatory capabilities are expandable by instruction, but it is currently unclear · how this will work.
- Means-ends analysis efficiency: Initial instruction p roduces a more elaborate network than is necessary. The Inst r uctor uses numerous intermediate goals, both to make his instruction sequences easier to generate and to allow complex procedures to be taught at all; miscommunication leads to a patchwork of var iant procedures; the PS uses a goal - encumbered monitoring mode. of operation insert ing supervisory goals and processes; etc. There are three . modes of varying efficiencies : a compiled, efficient mode; ordinary instruction mode; and the monitoring mode. Our primary concern is transforming from the ordinary form to the compiled form, while maintaining a capabil ity to revert back to the other two in debugging situations. This may not be attainable simply through instruction; architectural modif ication may be required.
- Utilizing distantly related knowledge: Knowledge about other tasks is imperfect for a given task; it is also embedded in methods and encoded in representations created for (and . local to) the distant task. All these aspects cause difficulty, even if contact is made (per 1). A clear symptom will be repetitive instruction to cover minor task variations. One approach is to avoid the difficulty by adopting uniform conventions for encoding. We think this won't work . . We favor attempting to map methods to methods (and representations to representations), using ideas from Merlin (Moore & Newell, 1973}.
Support. This. research was supported . in part by the Defense Advanced Research Projects Agency under Contract no. F44620-73-C-0074 and monitored by the Air Force Office of Scientific Research.
@inproceedings{ijcai1977p75,
title = {Problems in Building an Instructable Production System},
author = {Rychener, Michael D. and Forgy, Charles and Langley, Pat and McDermott, John P. and Newell, Allen and Ramakrishna, K.},
booktitle = {Proceedings of the Fifth International Joint Conference on
Artificial Intelligence, {IJCAI-77}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Raj Reddy},
pages = {337---1},
year = {1977},
note = {Knowledge Acquisition},
url = {https://ijcai.org/proceedings/1977/75},
}
TY - CPAPER TI - Problems in Building an Instructable Production System AU - Rychener, Michael D. AU - Forgy, Charles AU - Langley, Pat AU - McDermott, John P. AU - Newell, Allen AU - Ramakrishna, K. T2 - Proceedings of the Fifth International Joint Conference on Artificial Intelligence, IJCAI-77 PB - International Joint Conferences on Artificial Intelligence PY - 1977 SP - 337 EP - -1 UR - https://ijcai.org/proceedings/1977/75 ER -Copied
Machine-readable versions of this paper, for tools and AI agents:
Welcome
Future Conferences
AI Journal Division
IJCAI Proceedings
About IJCAI
IJCAI Awards