Group:
systems
Topic:
expert systems
Topic:
decision table
Topic:
knowledge representation by frames
| |
Summary
An artificial intelligence program can use heuristics and rules to limit the search space and select portions of the search space to explore. Heuristics may be manually or automatically developed. AM and Eurisko automatically generate heuristics to explore complex domains. They appear to work best when interleaved with manual pruning. This is effective in locating loopholes in the rules of complex games.
Heuristics must be used with care, otherwise they lead to unstructured programs that are difficult to analyze. For real applications, the quantity of specialized knowledge is immense. This can make heuristics difficult to use and develop. (cbb 4/98)
Subtopic: heuristics 
| Quote: another form of artificial intelligence uses heuristic or rule-base programming to solve a problem like a human seems to solves it
| | Quote: a heuristic is a contingent piece of guiding knowledge; limited scope [»lenaDB10_1982]
| | Quote: a heuristic domain must be observable with sufficient continuity and stability to make heuristics useful [»lenaDB10_1982]
|
Subtopic: automating heuristic discovery
| Quote: developing heuristics by analogy appears to be better than specialization and generalization; e.g., find examples before proofs [»lenaDB10_1982]
| | Quote: very few fields admit automatic data acquisition for developing heuristics [»lenaDB10_1982]
| | Quote: a maxim like 'compose two operations and study the result' generates too many possibilities; use an heuristic instead
| | Quote: automated discovery of concepts and heuristics most viable for large, unintuitive, unexplored search spaces [»lenaDB3_1983]
| | Quote: in automatic exploration of a semantic domain, brevity is the key attribute; useful concepts must be short expressions [»lenaDB3_1983]
| | Quote: AM and EURISKO do not mutate the meanings of concepts, they mutate the structural form of a concept in some representation [»lenaDB8_1984]
|
Subtopic: randomness in TD-Gammon
| Quote: TD-Gammon is successful because of the randomness of backgammon and a fairly smooth outcome function [»tesaG3_1995]
| | Quote: deep search can not be used for backgammon because there are several hundred possible combinations per ply [»tesaG3_1995]
| | Quote: even with a random initial network, TD-Gammon would terminate in, at most, several thousand moves
|
Subtopic: alpha-beta evaluation
| Quote: using functional programming, write alpha-beta evaluation as 'maximise . maptree static . prune 5 . gametree' [»hughJ_1984]
|
Subtopic: heuristics in AM
| Quote: for AM, started with a large number of concepts, slot definitions, an some initial values; only 1% of final knowledge added later [»lenaDB8_1984]
| | Quote: while most heuristic search programs use heuristics to prune, AM used heuristics to create the search space [»lenaDB10_1982]
| | Quote: in almost all cases, the discoveries made by AM were unexpected or unknown by the author; not obvious from AM's heuristics [»lenaDB10_1982]
| | Quote: AM selected a slot of some concept and worked to fill it; started with 100 concepts, about 20 slots for each (e.g., Examples, Conjectures) [»lenaDB10_1982]
| | Quote: to accomplish a task, AM obeyed relevant heuristics; they caused entries to be filled in and defined new concepts and tasks [»lenaDB10_1982]
| | Quote: almost half of the concepts proposed by AM's heuristics were interesting
| | Quote: each AM heuristic applied to a concept and all its specializations; so interpreter only considered potentially relevant heuristics [»lenaDB10_1982]
| | Quote: of AM's first 200 concepts, 124 were judged acceptable; of the next 300 concepts, only 29 were acceptable; AM started with 114 concepts [»lenaDB10_1982]
| | Quote: the heuristics used by AM clustered at the root and leaves of the concept specialization tree [»lenaDB10_1982, OK]
| | Quote: AM didn't discover mathematical concepts; it discovered small LISP predicates that happened to match math functions [»lenaDB3_1983]
| | Quote: AM exploited the natural tie between LISP and math by syntactically mutating small LISP programs [»lenaDB8_1984]
| | Quote: AM found many interesting results (not all known to Lenat) with multiple discoveries per heuristic, and many heuristics per discovery [»lenaDB8_1984]
|
Subtopic: heuristics in EURISKO
| Quote: EURISKO rests on not differentiating between meta-level heuristics about heuristics from object-level heuristics [»lenaDB10_1982]
| | Quote: EURISKO uses a large number of specialized slots; allows a condition to be an atom in some slot instead of bulky, if/then rules [»lenaDB10_1982]
| | Quote: EURISKO won the 1981 Traveler game; first attempt at a miniatures battle game (several hundred rules and constraints) [»lenaDB10_1982]
| | Quote: EURISKO uncovered anomalies, fortuitous rule interactions, and unrealistic loopholes; did not find rules about fleet and ship design [»lenaDB3_1983]
| | Quote: for the 1981 Traveler game, Lenat culled the heuristics generated by EURISKO every 12 hours or so; neither could have won alone [»lenaDB10_1982]
| | Quote: EURISKO quickly synthesized well-known electronic devices; but they were short sentences in the representation language [»lenaDB3_1983]
|
Subtopic: problems with heuristics
| Quote: heuristic or rule-base programming is dangerous because the rules are developed by trial and error; hard to predict effects [»parnDL12_1985]
| | Quote: shouldn't separate specific knowledge from general rules of inference; advice and knowledge should be linked together [»minsM_1981]
| | Quote: general AI methods failed when the specialized knowledge needed for real-world problems swamped their heuristic methods and data encodings
| Quote: a research program for AI is to hand-code a broad knowledge base, then acquire knowledge through reading, and finally learn by discovery
[»lenaDB1_1991]
|
Related Topics
Group: systems (17 topics, 530 quotes)
Topic: expert systems (8 items)
Topic: decision table (29 items)
Topic: knowledge representation by frames
(18 items)
|
