[LOCKSUIT]

What are Rewards for in a robot?

[ivan.moony]

Its purpose and purposes are what its father creator gives it.

khm... could be a mother too

but you're right, the question is completely opened.

[LOCKSUIT]

Rewards are for ranking the actions it generated.

Did you not know this?

[ivan.moony]

I've got a "working" AI system without rewards: it copies what it sees, but only those things that do not conflict with anyone. Pure learning from the scratch, but it has to learn enough before it knows what the conflict is. Then a do-not-conflict law gets injected from the outside.

The do-not-conflict law is the following: "do not do it if anyone suffers more than without doing it".

[LOCKSUIT]

What defines suffering is your pre-installation. When input triggers your reward, the actions are discovered.

[kei10]

The model of my True A.I roughly follows the nature of evolution of life. When it comes to life, it's always about one thing; Survival. Every patterns and system we see from a whale down to the scale of a amoeba, is all about survival -- either for oneself, or others. Which is why certain relationships between organisms can be understood.

This model consists of two major sections; The logic, and the emotion. The logic is the memory itself, whilst the emotion is the system that partially governs how the logic will flow, so as the logic also governs how the emotion will flow, where both must work as a pair together.

Through my research of the life evolution, there are three types of distinct patterns that can be observed; Going towards something, going away from something, and going nowhere. These patterns derives into the three major, observable emotion-feedback; positive, negative, and uncertainty. These three is nothing compared to other bigger patterns and systems that I found, that also must work together with one and another.

The reward, is nothing more than a part of the system that fuels/powers the trio-emotion-feedback, or more precisely, part of the positive-feedback mechanisms. Although suffice to say, without the reward system, it wouldn't make the emotion intelligence work correctly.

Therefore, no matter how much you babble about how awesome your solely -+ reward thing is -- it is incomplete, without other systems, it does not, and will not work completely.

[ivan.moony]

Suppose you are writing an intermediate program that creates an AI algorithm. The reward would be the time length being used until terminated. If you terminate the AI, it means something's wrong with it. Then you make another improved algorithm and run it again. The final goal for end-user AI would be to never be terminated.

[LOCKSUIT]

I never said in this thread that +/- is all that's needed.

And yes, when rewards reward senses, the label they receive may be neutral if it doesn't trigger a reward.

Ivan, do you understand that rewards are there to discover good and bad actions for a particular objective e.g. food and pain?

[kei10]

You've been showing us nothing but -+ reward, senses, ranking thing, and cortex thing.

[yotamarker]

in animals goals are preprogrammed (food, drink, thermoregulation)
for humans anything can be a goal

[LOCKSUIT]

I have how we can make up artificial rewards. We link the rewarded senses to the neutral sense. It won't be as thrilling though. This explains games/knowledge/movies/music. Plus you can keep changing them from good to bad ex. now if find the hat you lose.

[kei10]

Aahh... I finally realized what your A.I algorithm is missing...... You always focus only on those whatever you call "highest ranked" things. No wonder our viewpoint is always seems different, and I'm seeing yours half-wrong -- well, not half-wrong, but missing another crucial component. Well, never mind that, have fun with your rank thing.

[LOCKSUIT]

What crucial component?

There's stored senses & actions. Sense's rank & +/N/- LABEL, and strength & energy (can remember[faintly see] bright room proves energy differs from image brightness).

Selection of a sense - and therefore its linked actions, is by a match or self-ignition. A match is to the highest rank as long as similar (generalizes), + means do or speak "maybe" and even do a zero-action action!