[ CogSci Summaries home | UP | email ]

Hofstadter, D. R.& Mitchell, M. (1995). The copycat project: A model of mental fluidity and analogy-making. In Hofstadter, D. and the Fluid Analogies Research group, Fluid Concepts and Creative Analogies. Basic Books. Chapter 5: 205--267.

  ALTauthor = 	 {Douglas R. Hofstadter and Melanie Mitchell},
  ALTeditor = 	 {Douglas R. Hofstadter and the Fluid
Analogies Research group},
  title = 	 {Fluid Concepts and Creative Analogies},
  chapter = 	 {The copycat project: A model of
mental fluidity and analogy-making},
  publisher = 	 {Basic Books},
  year = 	 {1995},
  OPTpages = 	 {205--267},

Author of the summary: Jim Davies, 2002, jim@jimdavies.org

Cite this paper for:

[205] "Copycat is a computer program designed to discover insightful analogies, and to do so in a psychologically realistic way."
At a subcognitive but superneural level. [216]

Architecture is a hybrid of symbolic and connectionist.

[206] example from the problem domain: abc : abd :: ijk : ?
Where the agent is expected to do the same thing to ijk that was done to abc.

aabc : aabd :: ijkk : ?
If you answer ijll, then you are demonstrating fluidity and conceptual slippage, where the rightmost letter has slipped to the rightmost group of letters. Other kinds of things that can slip: leftmost with rightmost, successor and predecessor.

One might also see that aa and kk play similar "roles" in each of these. [207]

[208] "All the features of the Copycat architecturee were in fact designed with an eye to great generality...the Copycat project is not about simulating analogy-making per se, but about simulating the very crux of human cognition: fluid concepts."

"...even the most abstract and sophisticated mental acts deeply resemble perception." [210]

The Architecture has three major components [211]


[212] Each concept is a node. Links have different lengths, representing conceptual distances. There are roughly 60 concepts in this letter domain. Nodes have activation levels which can spread and decay. Each node also has a conceptual depth (e.g. opposite is deeper than successor, which is deeper than 'a'.) Roughly, it's a distance from directly perceivable things. Hard to notice, but when noticed, very significant. Deep concepts are resistant to slippage and decay more slowly.

[213] "The further away from direct perception, the more likely it is to be involved in what people consider to be the essence of the situation."

Because of this, unlike Gentner (1983), attributes (e.g. alphabetic-first) are stronger than relations (e.g. successor).

Links are labeled; the labels are concepts in themselves. The label's activation determines the length of the link. E.g. opposite links left and right.

Slippage is easier with shorter links.[214] The probabilistic cloud around a node is called a halo. [215] The net changes during the course of a run, but doesn't change permanently as a result. New concepts are not created.[216]


The workspace begins with a representation of the letters and which is rightmost and leftmost. [217]

An object's salience is a function of its importance and unhappiness. Importance is a function of activation level and the number of descriptions it has. Unhappiness is poor its links are to other concepts. Unhappy objects get more attention.

Objects can get linked by bonds, such as "sameness" for k and k. Bond have strengths based on activation, depth, and having similar bonds [218]. Objects linked with a uniform bond become chunked into groups. Groups also have desceriptions, salience, strength, and can bond, be scanned, and become parts of higher level groups.

Bridges are formed between two differnt frameworks (letterstrings, in this domain). This is an analogical map. Bridges have strengths, "depending on the ease of the slippages it entailed." [219]

Viewpoints are sets of mappings.


The coderack has agents called codelets. Scout codelets look for opportunities for effector codelets, which create and destroy things in the workspace. Created effector codelets are put in the coderack with an urgency value. [221] This value is a function of how well it fits into the workspace's viewpoint.

Botom up codlets notice all the time. Top down codelets look for higher level relations and can be called only by the slipnet. activated nodes can spawn them, as well as various "pressures" such as salience of workspace objects, and the activation, depth and stength of other nodes. .[223]

Bottom up codelets are constantly being added to the coderack, and codelets can call other codelets.,

[226] The workspace can only have one viewpoint at a time. Scouts can try to topple it, though, and bring about a different viewpoint.

[228] Temperature is a measure of how confident the workspace is with its viewpoint, inversely. Temperature controls the degree of randomness in decision making. Final temperature is a measure of how confident the system is in its answer.

[236]"...it is not intended that they shuld precisely reproduce the frequencies one would find if this problem were posed to humans, since, as we said earlier, the program is not meant to model all the domain0specific mechanisms people use in solving these letter-string problems."

Summary author's notes:

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
JimDavies (jim@jimdavies.org)
Last modified: Thu Aug 15 09:32:24 EDT 2002