[Ultron]

I did not really wish to post a new topic just for a simple, short question but it seems we do not have a more appropriate place for this type of things. I guess I'l use the chance to suggest a place (probably a topic) for quick questions regarding software / graphics. However, it would be inappropriate for questions regarding artificial intelligence in general since those are philosophical and are topics of discussion without strict answers, just opinions.


OK, the question.. How would a program (windows, c++) be able to monitor the system in a way that it can learn the consequences of each action (action -> reaction, function A cause status B)?


I have done similar things with an Arduino, but that is a far simpler device. Also, this requires monitoring a lot of data flow through many channels, that's about as far as my system knowledge can get me. I doubt anyone will come with actual technical knowledge so some ideas / concepts will be much appreciated.


On a side note, I believe nobody has done this before within the realm of computer science, however it is one of the essential principles of intelligence in nature.

[Don Patrick]

Depends on what exactly you want to monitor and what exactly the actions are. Mouse clicks? Keyboard presses? Window titles? Graphical changes? All those can be logged with ease (see task automation), and one might assume that one event causes the other if within a narrow consequtive timeframe.
What would you have the program do? Randomly click on your screen until something happens? The difficulty would be for the computer to know when it has achieved something worthwhile, or it may become very proud of its ability to open up 1000 browser windows. It seems like that would be a task for goal-set evolutionary algorithms.

[ivan.moony]

I believe that it could be achieved by computer imagination and induction.

First, you blindly construct whatever theory in a form of A -> B. Then you inductively check the memory of events to see if the theory holds for all relevant events (i.e. Did B came after every A in the past). If the theory holds, you remember it and use it also for genetic steering of the future guesses from the beginning of the story. Note that in the future an event can pop up that makes the theory false, as induction you do is mostly done on incomplete number of examplars.

At the first, it is a brute force of checking whatever theory can be constructed by random. Later, when you get enough hits, you have a case to use parts of positive guesses selected by genetic algorithm to get more positive results in the less cycles.

There is an alternative to a pure random construct: generalization. You take two or more cases from the past and you construct an expression that describes all those pickups in one sentence (like from 1+2=3 you can try integer+integer=integer). Then again, induction check gives you an answer of correctness.

But to implement all of these you have to have a language expressive enough to describe the matter you want to analyze. I'm working on such a language.

[Ultron]

I thank you both! You have inspired my imagination on different approaches towards this issue, however the goal remains that the program must be as simple and as effective as possible.


The goal is NOT 100% accuracy - sometimes we observe a man saying several words to a handsome woman on the bar and she accepts the offer for communication. This is based on audiovisual input and a simple conclusion that A (the pick-up line) lead to -> B (the lady joins the conversation).
After the robot attempts the same procedure, he fails. A did not lead to -\> B, because it did not match the value for a specific variable - the robot was not a handsome man. But the robot also did not know that this variable was part of the equation.


The problem is, while I could theoretically (and inefficiently) log all variables within a computer system, that is inconceivable in life because there are far too many variables and possible values. The question is, how do humans do it? Okay, once we see that A does not lead to B, there must be more variables to A or different values to the existing ones. We start searching for others - we monitor the successful example again but we use some complex psychological constructs to identify other variables (surely we won't log everything else, such as the wattage on the lightbulb shining over the bar - but in other cases that may be exactly the key).


That's the construct / algorithm I am interested in - as part of a theory. The actual goal and original question related to a far simpler issue contained within a simple computer system.
According to the arrogant community of StackOverflow, this task is nearly impossible and would require one to create an operating system of his own. This however, only fuels my persistence.

[8pla.net]

Program observing system behavior

Unknown to you, but between 1 to 8.

Guess what is unknown to you... 5
No. That is more than what is unknown to you.

Guess what is unknown to you... 3
No. That is more than what is unknown to you.

Guess what is unknown to you... 7
No. That is more than what is unknown to you.

Guess what is unknown to you... 1
No. That is less than what is unknown to you.

Guess what is unknown to you... 3
No. That is more than what is unknown to you.

Guess what is unknown to you... 4
No. That is more than what is unknown to you.

Guess what is unknown to you... 2
Yes! The unknown is now known to you.


References:
http://www.cplusplus.com/reference/cstdlib/rand/
http://www.cplusplus.com/doc/tutorial/functions/

[Ultron]

Maybe it's me not sleeping for two days, or maybe I really can't see what method / principle you are trying to point me towards. Clarify.