Intelligent Systems And Their Societies

Walter Fritz

español            

The author has added this page to the e-book, since it may be of interest and of help to others working in the same direction. He is presently constructing the General Learner 3 (GL 3).

The GL 3 is an artificial brain that not only determines its external actions by rules but also performs the internal processing of incoming information and the selection of an external action by internal rules. The GL3 itself can modify and possibly improve external and internal rules while it is running.

• The C language programmer will find it strange, that the GL3 program modifies itself while running.
  All that happens is that the part of the program that is the central brain runs elementary C language functions. These do NOT change. The input to these function is by concepts, and these are data and DO change. Rules determine which elementary C language function is done in a given moment, and these rules are data and DO change.
  In this way the program changes its functioning, by learning new and better rules.

Theoretical note about intelligence
In a most general sense we can say that the activity of a brain, biological or artificial, is the discovery of patterns, the storing of patterns and the use of patterns.

Learning, in the artificial brain, is basically a discovery of patterns. It is a finding of patterns within the input data (senses) and also within the output data (acting). These patterns are newly created concepts.
Furthermore learning is a finding of patterns between the input and the output (These are newly created response rules). Rules are similar to "productions" used in expert systems. They also have an input and show the corresponding output. The difference of rules from productions is that rules do not have the input and output stated in words, but in concepts, and that these can be learned by the program itself.
The artificial brain uses patterns (response rules) when it decides what to do in a given situation.

• It seems to me that in the biological brain, dendrite connections are the pattern. The brain continuously tries to create further useful dendrite connections and inactivates those connections found useless. Making an axon connection to another neuron nearer the output of the brain is somewhat like making a response rule.

The previous General Learner
The previous General Learner took the sensations and then looked up rules of action, that were appropriate to the sensations. This process was done by the program, using C language functions. As we will see, in the General Learner 3 this process is performed by rules.

• The authors artificial brains are computer programs, written in the "C programming language". Parts of these programs, namely the sub - programs, are called functions in the C language.

The General Learner program had senses (the input), a "central brain" (the main cognitive architecture) and actuators (the output). The senses get information from the environment and send them to the "central brain" as elementary sensations. The actuators get elementary actions from the "central brain" and perform them.

The "central brain", part of the computer program, has no way of knowing what the input bytes and output bytes mean (for human beings). Its main objective is pleasing the person operating the computer. The person expresses approval and disapproval by the up and down arrow keys.
The brain continuously observes its environment and what actions are done by itself and the person. It looks for patterns between parts of the environment and these actions and remembers them as rules. Also it can learn abstract concepts such as "vertical", if we draw examples of horizontal, inclined and vertical lines on the computer screen.
See details in General Learner (For continuos reading, like a book - do not enter here now)..

What is new in the General Learner 3 ? (GL 3 )
Previously, only actions that are convenient in a certain situation, were learned. Now, in addition to this, thinking is learned, that is, better thought processes in the "central brain" are learned. Here "better" means resulting more often in adequate actions.
Previously the functions of the "central brain" was realized by C language functions. Now these functions are realized by rules and details of these rules can be modified by the General Learner itself.

• We can say that, in the artificial brain, the hardware (the sense organs), receives sensations. The software (the program), presents these sensations to the mindware. The mindware (the concepts and rules of the "central brain"), determine what actions to do. The software receives the action concepts and orders the hardware to do the actions.

In the GL 3 a response rule has several parts. One part is the "sit1" the present situation, to which the rule is applicable. Another is the future resulting situation, the "FS" which will be obtained by the application of the rule. This is the objective or goal of the rule.
The part, that was previously called the action part, now is the "IntSit" part. Namely the intermediate situation that has to be reached by lower level rules, in order to obtain the FS. All of these parts are expressed with concepts.

Even at the start, there is no fixed rule structure. The GL 3 builds up the initial present situation. Also it has a future objective situation, both expressed with concepts. The GL 3 looks up rules that satisfy the present situation and the future situation and executes them.
Both rules and concepts have a value attached to them, that increases if they are part of a successful rule. Elementary sensations and elementary actions now are called external elementary rules, not elementary concepts (Because they really are activities).

From the start, the GL 3 has a set of internal elementary rules supplied by the programmer. Each internal elementary rule has an empty "IntSit" part. So elementary rules include the sit1 to which it is applicable, an empty IntSit and the FS (the resulting future situation).

Approval and disapproval
The activity of handling the approval or disapproval given by the person remains a C language function and cannot be modified by the brain of the GL 3. This is most important, since otherwise control over the main objective, and with that of the actions of the program, is lost. The C program function gets the evaluation from the person, and adjust all recently used rules, internal and external, accordingly.

Possible Applications
Chatting with a person in English and playing board games on the screen. Future applications could include translating English to Spanish, functioning as a robot mind, or a C language function writer if the correct input and output is given.

For more details see the General Learner 3 (Enter for continuos reading, like a book)..

Exactly how can the GL 3 learn better rules
for its internal processing?
External rules, the rules that respond to an external situation, indicate the action to be done in the environment. These rules have been learned from experience by the prior GL. Also the prior GL has abstracted and generalized these rules, to get more generally applicable rules.

On the other hand internal rules are those rules that do the internal processing in the artificial brain. These are rules that respond to an internal situation and try to get to an internal future situation. The prior GL could learn external rules, but had no internal rules. The job of the internal rules was done by parts of the C language program.

The differentiation between external and internal rules is somewhat arbitrary, since both have the same shape and are used in the same way, but not for the same purpose.

It seems that the learning of internal rules should be just the same as the learning of external rules. Here the program should observe the internal situation, the required objective situation and the action performed. From these it should build up new internal rules.

• For instance there could be an internal blackboard where situations and actions are recorded. The "blackboard" is an area of memory in the computer into which both the GL 3 and the person can write. By clicking a menu it is represented on the screen. The person could write on this blackboard and the GL 3 observes, and learns, from this activity.

Also the GL 3 has an initial set of internal elementary rules. It looks for two rules where the FS of one is identical to the sit1 of another, and combines them into a new rule.

It learns from the effect of these rules. Further, as in the previous GL, the GL 3 could abstract and generalize, thus creating further internal rules. The main point seems to be that the GL 3 builds up its rules from lower level rules and initially from elementary rules.

See details in GL 3 Learn rules (Enter for continuos reading, like a book)..

Comparison with Neural Nets
Since each response rule performs the function of a neural net, The GL 3 has the equivalent of 20 000 interconnected neural nets. Each rule receives inputs and gives outputs to other rules. This is equivalent to 20 000 neural nets receiving inputs and giving outputs to each other.
The GL3 creates rules as they become necessary, that is the equivalent of creating new neural nets in this great web of interconnected neural nets.

Final Remarks and present state
The above is a plan or specification of the GL 3 which is under construction. Usually, when such a plan is executed, changes will prove necessary. That probably is also the case with the GL 3.
It remains to be seen, if the GL 3, by producing new rules, can really improve the overall performance of the system. It looks to me that this is likely, but that can only be shown by building the system, experimenting with it and modifying it.

At present the GL-3 is runing nicely in the awake mode (exterior activity). The sleep mode (interior review of the memory and build up of new rules) is about 20% built. The program builds up elementary rules nicely and uses them to perform exterior actions. But the elementary rules are stilll too "makro" and need to be decomposed more.

It is quite important and unusual that a program can learn new methods of "thinking"; new methods of processing the incoming information and choosing an adequate action.

For continuos reading, like a book - continue here.
Jump to the top of this document /A.I. Programs, that are complete intelligent systems /Robby, A Household Robot /E-Book Contents.