Memory retrieval

    Consider the most basic type of understanding--a reasoner is presented with a concept which they already possess in memory. For understanding to occur, the full concept must be retrieved from memory. This will then allow       prediction, abduction, and explanation to take place. Because of this fact, memory retrieval is the first task required by the   reasoning supertask. As concepts enter this supertask, a retrieval request is sent to   memory. memory returns to the reasoning supertask a set of concepts which represent the ones determined to be most similar to what was requested.

  

Figure 24: Memory retrieval algorithm

The high-level algorithm for memory retrieval is shown in Figure 24. The cue set represents a collection of attribute names and values which the reasoner would like to have the retrieved concepts contain. These form a retrieval request. For example, the request ``retrieve all the red-colored octagon-shaped objects you know'' is interested in the following aspects of concepts:

1.

The concepts which are returned must be objects rather than actions or states

2.

The concepts must possess the attributes for color and shape

3.

The color attribute should be filled with the value of red

4.

The shape attribute should be filled with the value of octagonal

The indexed retrieval simply refers to the fact that concepts are retrieval based on known attributes. Memory should be searched for concepts which match these attributes and these values. However, match is actually not strongly defined. For example, what if the memory possesses a concept which is a maroon-colored, octagonal object? Should this be returned? Maroon and red are related, after all. Suppose the memory contains many octagonal objects with unknown colors? What about many red objects of unknown shape? And so on.

This discussion is meant to illustrate that there is more than a ``right vs. wrong'' outcome to the memory retrieval request. First, an exact match may be discovered in the time allowed for the memory retrieval. If this occurs, then the concept being understood has appeared to the reasoner before, in the exact form it now has. This is a kind of understanding which I refer to as pure recognitional understanding. Once the concept is recognized, the reasoner can use prior knowledge about the concept to make       predictions, explanations, and abductions about it. Closely related to this is what happens when the exact concept cannot be retrieved, but concepts which match on the   primary attributes are returned. This is an example of an instantiationally novel (I-novel) concept being understood. Since the   secondary attributes do not affect the function of the concept, the understanding process continues as with the pure recognitional form.

If an exact match is found, then the concept being considered is a known one, and understanding has occurred. It might also be the case that only similar objects are returned; no exact matches exist. Finally, it might be the case that some concepts are returned which were judged in some way to be similar to the concept in question but which do not provide enough of a match for   reasoning to be satisfied. For example, a match may be returned by the memory retrieval task which does not provide accurate predictions for future events. In a case like this, the retrieval task has failed.