Ais Development

It's time to develop true Artificial Intelligence System, see the overview at http://www.dennisgorelik.com/ai/overview.htm

I invite you to take part in AI development http://www.dennisgorelik.com/ai/YourHelpInPCnousDevelopment.htm

Discussion:

I'm afraid that I have to disagree with AIS being the right time now. Even though pc's may be strong enough and even though we know how to simulate neurons, that has never been the hardest part in AI. Having studied it and tried it myself I always found that the learning function is the hardest to get at. How do you tell what is right and what is wrong to a computer? Humans learn this slowly over the course of years, and in the beginning they might hurt themselves or so if something goes wrong, how do you tell a computer something's wrong?By giving it a false at the output? All right...how do you come up with those TRILLIONS of training patterns? -- ChristophePoucet?

How will AIS (intelligent computer) know what is right and what is wrong?

AIS has several "basic" goals. These goals evaluate consequences of AIS actions. Some consequences satisfy the goals, other consequences dissatisfy the goals. AIS use the evaluations to update desirability levels of "neurons responsible for the actions".

Afterwards, when AIS will make decision, AIS will just choose the most desirable concept.

So, AIS is guided by the "basic" goals.

How do you tell what is right and what is wrong to a computer?

You may affect AIS behavior by three ways:

  1. Program some of AIS goals. For instance: "Read", "Talk", "Think", "Find correlations among concepts", "enrich bank account" goals.
  2. Work as a goal. You may observe consequences of AIS actions. Then evaluate them. For this purpose AIS has to have "Get evaluation by the owner" goal.
  3. Talk with AIS and try to convince AIS in something. This way you usually use if you want to affect human's behavior.

DennisGorelik (Refactored to this page by ChristophePoucet?)

I'm sorry but what you state above you're being -very- general. Having tried AI development myself, including neural networks and genetic algorithms, I know from experience that once you start to implement those basic goals, you will find they are not basic at all.

Problem 1: Knowing which neuron helped accomplish the task is VERY hard, especially if you only got as response 'good' or 'bad'. BackPropagation? might work, but it will take ENDLESS cycles to actually get somewhere.

Problem 2: You might get stuck in a local maxima that by no means approaches the AI you're aiming for.

Problem 3: As for finding out when a basic goal has been achieved. Even this is very hard. For example, let's say a cart has to drive from point A to point B. How will you know once you arrive at point B and get the good cart response, which actions led to you arriving to point B? How do you know which ones to reward or not? It could be that the cart took a bad turn somewhere in the middle but rectified in the end.

-- ChristophePoucet?

Christophe, I put my answer here: 

http://www.dennisgorelik.com/ai/QuestionsAndAnswers.htm
You are welcome to copy it to Wiki (but I have no idea what will you do with references).

See also: AisDevelopmentLinks

Refactoring notes

My recommendation about AisDevelopment refactoring:

  1. Big page should be divided into small parts. I think that several small pages are better than one big page because:
    • Small page is easy to read.
    • Text on small page has better cohesiveness.
    • Small page is not boring.
  2. You are welcome to refactor outer links (E.g.: http://www.dennisgorelik.com/ai/):
    • Copy outer text to a Wiki page.
    • Modify it.
    • change reference to the Wiki page.

-- DennisGorelik

EditText of this page (last edited May 25, 2004) or FindPage with title or text search