Learning search strategies through discrimination in Int. J. Man-Mach. Stud. 18,6 (June 1983), 513-541;
|
 |
|
Post a Comment
|
|
|
|
|
CONTRIBUTORS:
|
|
|
JOURNAL:
|
Computing Reviews ,
25(6),
289 -
290.
|
|
|
|
YEAR:
|
1984
|
|
PUB TYPE:
|
Book Review
|
|
SUBJECT(S):
|
GENERAL TERMS: HUMAN FACTORS. PERFORMANCE THEORY
|
|
DISCIPLINE:
|
Computer Science
|
|
HTTP:
|
|
|
LANGUAGE:
|
English
|
|
PUB ID:
|
103-426-836
(Last edited on
2006/05/23 12:17:51 GMT-6)
|
|
ABSTRACT:
Learning search strategies through discrimination. Int. J. Man-Mach. Stud. 18,6 (June 1983),513-541.
The author describes SAGE (Strategy Acquisition Governed by Experimentation) as a system in the artificial intelligence domain where a computer program solves a problem by using a generalized theory or rules (strategy) and various sequences of experiences which usually include errors. Problem solution causes the system, which is nonhuman, to learn.
It is stated that few researchers have attempted to construct strategy learning systems until recently. Basic principles for such learning systems have been proposed by Langley and Simon [I] and include generation of alternatives, knowledge of results, causal attribution, and hindsight. Using these principles as criteria, five recently reported problem solving strategies are examined.
ALEX, developed in 1978 by Neves [2], is identified as an adaptive production system which incorporates two of the four basic principles. However, ALEX does not possess a scheme for the knowledge of results and is unable to recover from an error if an incorrect rule is learned.
Anzai [3,4,5], in 1978, reported on a system which improves its strategies when solving the Tower of Hanoi problem. While it does incorporate rule-learning behavior and qualifies as satisfying all four of the Langley and Simon criteria, it does have several undesirable features. Its task domain is well-known and specific, but not a substitute for a generalized program. Also, the design assumes that all learned rules are correct and there is no need for recovery from error if an incorrect rule was constructed in the process.
Brazdil [6] described ELM in 1978 as a PROLOG program based on symbolic logic. It has a set of operators for problem solving, but the order is unknown. After solving a problem, it knows the successful rule or rules, giving them a priority for subsequent problem solving efforts. It is capable of reassigning priorities after some experience with various problems. Thus, ELM can determine when performance is improved, assigns credit and blame to system operators, and modifies its behavior based on this knowledge. However, ELM requires a benevolent tutor for sample solutions and does not generate behaviors by itself.
In 1981-82, LEX was described by Mitchell, Utgof, Nudel, and Banerji [7] as a strategy learning system incorporating all of the four Langley and Simon basic principles. While there are disadvantages, these may be viewed as minor. Langley sees LEX as having considerable potential.
Finally, Neches [8], in 1981, described HPM as an adaptive production system with learning heuristics stated as condition-action rules. Langley views HPM as incorporating all four of the criteria, but "in rather unusual ways."
Using these examples of previous research, the author concludes that while the findings reveal important evidence of problem solving strategies, none of the systems has shown that it can learn in more than a single domain. In essence, Langley adds a fifth basic principle: the system must be domain-independent and learn on a number of different tasks. The SAGE system is purported to meet all five basic principles.
The paper, while a valuable contribution to AI research, does not mention the time parameter; it would have been helpful if the time-to-produce-solution was included. Also, PRISM is an attractive language but the author does not describe the hardware system which runs it, nor the memory requirements for a typical problem solving event. Practitioners would be interested in this information. The paper is recommended to AIers who wish to keep up with the flood of professional papers now reaching publication.
|
|
|
|
STATISTICS
|
|
Click on # to view
|
|
Citations
|
|
0
|
|
References
|
|
0
|
|
Comments
|
|
0
|
|
Quality
|
|
0/0.00
|
|
Interest
|
|
0/0.00
|
|
View(er)s
|
|
2/97
|
|
|
|
|
|
|
| Prev |
Next |
|