Examples of analogical reasoning were presented in Section 7.3.2. Describe an appropriate

representation and search strategy that would allow for identification of the best

answer for this type problem. Create two more example analogies that would work with

your proposed representation. Could the two following examples be solved?

a. hot is to cold as tall is to {wall, short, wet, hold}

b. bear is to pig as chair is to {foot, table, coffee, strawberry}

7.3.2 Copycat

An often-heard criticism of traditional AI representation schemes is that they are static and cannot possibly reflect the dynamic nature of thought processes and intelligence. When a human perceives a new situation, for example, he or she is often struck by relationships with already known or analogous situations. In fact, it is often noted that human perception is both bottom up, that is, stimulated by new patterns in the environment, and top down, mediated by what the agent expects to perceive.

Copycat is a problem-solving architecture built by Melanie Mitchell (1993) as a PhD dissertation under Douglas Hofstadter (1995) at Indiana University. Copycat builds on many of the representational techniques that preceded it, including blackboards, Section 6.3, semantic networks, Section 7.2, connectionist networks, Chapter 11, and classifier systems (Holland 1986). It also follows Brooks approach to problem solving as active intervention in a problem domain. In contrast with Brooks and connectionist networks, however, copycat requires a global state to be part of the problem solver. Secondly, rep- resentation is an evolving feature of that state. Copycat supports a semantic network-like mechanism that grows and changes with continuing experience within its environment.

The original problem domain for copycat was the perception and building of simple analogies. In that sense it is building on earlier work by Evans (1968) and Reitman (1965). Examples of this domain are completing the patterns: hot is to cold as tall is to {wall, short, wet, hold} or bear is to pig as chair is to {foot, table, coffee, strawberry}. Copy- cat also worked to discover appropriate completions for alphabetic string patterns such as: abc is to abd as ijk is to ? or again, abc is to abd as iijjkk is to ?, Figure 7.27.

Copycat is made up of three components, the workspace, the slipnet, and the coder- ack. These three are mediated in their interactions by a temperature measure. The temper- ature captures the degree of perceptual organization in the system, and controls the degree of randomness used in making decisions. Higher temperatures reflect the fact that there is little information on which to base decisions and thus they are more random. A tempera- ture drop indicates the system is building consensus and a low temperature indicates that an answer is emerging and reflects the programs confidence in that solution.

The workspace is a global structure, similar to the blackboard of Section 6.3, for cre- ating structures that the other components of the system can inspect. In this sense it is also much like the message area in Hollands (1986) classifier system. The workspace is where perceptual structures are built hierarchically on top of the input (the three strings of alpha- betic symbols of Figure 7.27) and gives possible states for the workspace, with bonds (the arrows) built between related components of the strings.

The slipnet reflects the network of concepts or potential associations for the compo- nents of the analogy. One view of the slipnet is as a dynamically deformable semantic net- work, each of whose nodes has an activation level. Links in the network can be labeled by other nodes. Based on the activation level of the labelling nodes, the linked nodes grow or shrink. In this way the system changes the degree of association between the nodes as a function of context. The spreading activation among nodes is encouraged between nodes that (in the current context) are more closely related.

The coderack is a priority based probabilistic queue containing codelets. Codelets are small pieces of executable code designed to interact with the objects in the workspace and to attempt to further some small part of the evolving solution, or, more simply, to explore different facets of the problem space. Again, the codelets are very much like the individual classifiers of Hollands (1986) system.