Intelligent Systems And Their Societies Walter Fritz

Storing and Generalizing and Forgetting


Storing the Response Rule
The response rule, as it was finally applied, may be quite different from the response rule that was originally selected. It is the applied response rule, however, that the GL stores in memory and in the chronological file. The GL uses this file to create general response rules during a period of external inactivity.

Also, a person may act without intervention of the computer. In this case the computer learns by copying the actions of the person. It takes the present situation and adds to it the action that the person did, and thus creates a new response rule. If the person does a series of actions, the GL creates a series of response rules for later use.


Plan Making
In other artificial ISs, a variety of quite elaborate plan making procedures have been built and explored. Some of these proceeded by using response rules to move forward from the given present situation to the objective. Others proceeded by working backward from the objective situation to obtain the present situation. Still others proceeded by working in both forwards and backwards directions and trying to make these paths meet at some middle point. See also Mental Methods (For continuous reading, like a book - do not enter here now). in Intelligent Systems.


Creating Generalized Response Rules
At the start of the run, the system does not have any response rules in storage. As GL gathers experiences it begins to create response rules, as we have seen.

But these response rules are usable only if the present situation, or part of it, corresponds exactly to the stimulus part of a response rule. Normally, this is not the case. The environment of an IS varies far too much. Even in our simple computer environment of key presses and lines on the screen, the number of possible variations is enormous. Thus, in order to improve its abilities, it is necessary that the GL creates generally applicable response rules. But reviewing the existing memory and creating new response rules, concepts, and rule patterns is very time consuming; in fact, it would interrupt the ability of GL to smoothly control its external activities. Therefore, the GL does this internal review during a period of external inactivity. (We call it a sleep period.)

The mechanism that the GL uses to create generalized response rules begins with a review of the new experiences that it has cataloged. This review always starts with the last response rule in the chronological file. After the review, the GL deletes this response rule and moves up to the new last rule. At the end of the sleep period the chronological file is empty and a list of new, generalized response rules has been created.

The review of a particular new experience goes like this: First, the GL takes each concept of the stimulus side of the last response rule and puts all those response rules that also have this concept in their stimulus side into a list. To find these rules fast, it adds the response rules (the numbers) stored in the pAddrRr branch of the concept to the list. Then it eliminates duplications, response rules with only negative weights, and those containing rule patterns. This produces a list of response rules that have something in common in their stimulus side.

Then GL analyzes the list. It uses two basic manners of review: vertical and horizontal comparisons.

Vertical Comparisons
For a list of response rules in memory, the GL compares the situation (or response) vertically (situation to situation, or response to response) between the response rules. By this comparison it can build up concepts from concrete examples or from parts , as well as new response rules based upon generalizations or symmetry.

More Abstract Concepts.  If the GL finds response rules with all parts identical except one concept, it creates and stores a new concept that has the two different concepts in its link to concrete examples. This can be the case in the stimulus part or the response part of the response rule. Further, it creates and stores a new response rule that contains the new concept in place of the concrete concept.

For instance:
draw me a curve -> (a curve is drawn)
draw me a arc -> (the same curve is drawn)
Here "curve" and "arc" are concretes of a new concept, namely a certain line that is bent.
At the same time the GL creates the corresponding generalized response rule using the newly created concept.

Create Concepts from parts.  Similarly the brain creates a new composite concept out of parts. This reduces greatly the amount of information it has to handle at each mental process. This is of special advantage in reducing the amount of elementary concepts.

Suppose this vertical comparison process finds two response rules where a series of concepts in the situation or response part are identical. To reduce complexity, the GL here replaces the series of concepts by a new composite concept (e.g. "A line" would be a composite concept, composed of "A" and "line"). Then it stores the composite concept and creates a new response rule containing the composite concept in place of the series of concepts.

Generalizations.  If there are various response rules that are identical, except that some have more concepts in their situation, the GL considers them superfluous and forms a generalized response rule with only the remaining concepts.

For instance (the person types and draws):
"draw a long line" -> (drawing of a line)
"draw me a line" -> (drawing of a line)
Here the GL creates the generalized response rule with the concepts common to both:
"draw a line" -> (drawing of a line)

Symmetry.  If the GL finds response rules that have drawings in both the stimulus and response part, then it creates the corresponding mirror response rules, interchanging plus x for minus x and plus y for minus y. (This change is often used in games).

Horizontal Comparisons     
For a list of response rules in memory, the brain compares concepts of the situation horizontally with concepts of the corresponding response. This horizontal comparison generates rule patterns.

The system stores the type of rule pattern and information about where in the response the rule pattern is to be applied; also to what concept or amount of concepts, and at what depth of the concept (to which detail of the concept) it is to be applied.

While many comparisons are possible; of these, we have used only four up to now:

  1. Same Concept
  2. Transformation
  3. Repetitions
  4. Relationships

1. Same Concept.  The first case is when it is found that the same concept occurs in both the stimulus and the response of a response rule. If it is found in more than 70% of the cases that the concept is in the same position in all of the stimulus sides and that it is in the same position in all of the response sides, then the GL creates a rule pattern. The position in the stimulus may be different from that in the response, but all positions in the stimulus have to be the same, as well as all positions in the response (first, second, last, next to last and so on).

Example:  Suppose that the input (situation) is "I am Peter" and the output (response) "Hello Peter" and this is repeated with other names. Then GL will create a rule pattern and a response rule: "I am (rule pattern)" > "Hello (rule pattern)".

2. Transformations.  The second case is when a certain concept in the stimulus indicates that a certain different concept (or part of the concept), occurs in the response. If this is the case in more than 70% of the response rules that contain the first concept, the GL creates a rule pattern. These rule patterns indicate thin or fat lines, horizontal or vertical figures, inclined, short or long lines and so on.

Example:  For instance, the concept related to the word "vertical" in the situation may have a rule pattern with the values of x within the concept of the response (very low values of x).

3. Repetitions.  The third case is applicable when the same concept in the stimulus corresponds, in more than 70% of the cases, to a certain amount of repetitions of a concept in the response. (This would be a concept designating a number.)

Example:  For instance in "draw me three circles" the concept for "three" is related with the amount of identical concepts in the response of the response rule.

4. Relationships.  This comparison looks for the occurrence of a certain concept in the stimulus side that relates rule patterns of the third type; namely rule patterns related to amounts. If, in more than 70% of the cases a certain concept of the stimulus side of a response rule is surrounded by two concepts that have rule patterns of type three, and if the repetitions of the response side have a simple arithmetic relationship between the rule patterns of the stimulus side, then the GL creates a rule pattern. Thus, this is a rule pattern that designates a relationship between rule patterns: in this case a relation of addition, subtraction, division or multiplication.

Example:  For instance, the stimulus side "draw two times three lines" fits this pattern.


We have found that even with the large RAM memories that exist in today's personal computers, memory fills up too fast. This difficulty can be alleviated if we allow the artificial IS to forget. The method GL used is that it forgets those response rules and rule patterns that it has not used for the longest time. It also forgets concepts that are no longer in use in any response rule. This activity also occurs during the sleep period. In a previous system we used a method that was based on forgetting the "least important" response rules. It is believed, though, that further experimentation is necessary to find the best method.

For continuous reading, like a book - continue here
Jump to the top of this document /General Learner /Artificial IS /e-book Contents.

Last Edited 11 April 2013 / Walter Fritz
Copyright © New Horizons Press