Ai Forums Home Welcome Guest    Wednesday, September 20, 2017
Ai Site > Ai Forums > The Artificial Intelligence Forum > Proper maximalisation by implementing correct feedback Last PostsLoginRegisterWhy Register
Topic: Proper maximalisation by implementing correct feedback

absk
posted 4/24/2016  21:58Send e-mail to userReply with quote
The following may seem weird, but actually I put quitte a lot of thought into it so looking forward for some reactions.

Two theories:
- All systems are steered by some kind of maximalisation mechanism.
- Humans steer by maximalizing happiness.
- A society (should) try to maximise the collective hapiness.
- Companies steer by profit maximalisation.
- Many problems arise from a incorrect feedback mechanism. Or the former maximalisation prosess is hindered by incorrect feeback mechanisms

Intelligence.

We could also argue that true intelligence is reached by emplying sufficient feedback.

- Reacting intuitively or in a reflex can be seen as reacting directly on an input. I.e. an input is directly translated into a response. The signal goes straight from input to output.

- Reasoning however uses the information from this first pass for further analysis.

In the case of Deepmind AlphaGo, the intuitive part is done by the policy neural network which specifies a number of suitable moves. Further 'reasoning' is provided by a tree seach algorithm (also employing neuralnetworks).


It is to be expected that in the human brain that this further reasoning involves some kind of neural loop. Information from the first intuitive pass is looped back and further analysis is performed. For example:

Say in a picture A cat is blocked by a dog. The dog neuron is happy but the cat neuron is only half happy. In the next iteration it is identified that indeed where the rest of the cat should be there is a dog and therefore that the dog overlaps the cat. In the end the "is "blocked" and dog neurons are completely happy while the cat neuron is either half or wholy happy. Happyness is therefore maximalised.


No feedback present; I.e. if an action does not have a consequence this can be problematic.

Example 1: CO2 emission is (almost) free of charge. So it doesn't matter whether you emit 1 or 2 tons of CO2 in both instances you don't have to pay a penalty.
Solution: Implement the carbon tax

Example 2: Companies are inefficient and problematic entities. The owner of a company does not have to listen to it's employees. If he makes mistakes he still will be the owner. Therefore there again is no correcting mechanism to punish mistakes (besides him running the company into the ground). Information will mainly go from the top down to the lower ranks. Feedback from the lower ranks is often not welcomed. Wages are often not decided by performance but by politics. The owner will try to maximize profit while minimizing costs. Therefore wages will be kept as low as possible, there is no internal mechanism which ensure the employees get a fair salary.
Solution: Some kind of new entity? For example a foundation which is allowed to make a profit and therefore more or less operates like a company. Profit maximalisation will still steer for efficiency but earnings are now (potentially) fair devided and positions are now (potentially) fair awarded.
Position, wage, hiring and firing of people should be self organizing and needs to be based on performance. The people are best suited to evaluate the people directly above and below them. Therefore the people from the second level are chosen from the ranks of the first (lowest) level. In other words the leadership is chosen from the bottom up.
Etc... Start capital could be provided by the state.

Example 3: In current democracies a problem is that the top leadership is chosen by everybody. Therefore the amount of education, intelligence or experience has no influence on the voting process. Therefore the current democracy at the best of times cannot make more intelligent decisions than the common man on the street. While at the worst of times the decision is a compromise between the two or more governing political factions. Whith an infinitesimal amount of political factions this would more or less equal some kind of average noise. Therefore not providing adequate feedback to steer the country into the right direction.


Positive feeback: Pushing system out of equilibrium into another state.

Example 3: The (atomic) bomb is a example of positive feedback. Positive feedback will push things out of equilibrium into another state. Often this is not desired.

Solution: Do not make and or detonate them...


Negative feedback: Maintain system in current state.

Example 4: A certain company will lobby for it's interests. Due to political lobby the position of the company becomes better and it becomes more difficult for competitors. Therefore change will be resisted resulting in a very conservative industry.

Solution: Forbid paid lobbyist will level the playing field for innovative startups and promote change and innovation.





keghn
posted 4/24/2016  23:00Send e-mail to userReply with quote
Ya. I think about that allot too. In a little different way.
From primitive level to a very complex level based around energy use
of a system.


absk
posted 4/25/2016  02:24Send e-mail to userReply with quote
Another thought is that for a very simple stimulus response model. Like the Deepmind Atari game playing programs. We only have two entities, namely the World (W) and the intuitively creature (I). Therefore the world is perceived by the creature and it immediately reacts.
However at Deepmind this however does not work well for the game PacMan since a sense of the world is missing, or in other words the walls are in the way...


A more complex creature would add a third component namely a imaginary model of the world (IW).

This more complex creature would still perceive the world. Come up with an intuitive response but next does not immediately output this response but gives this as an output to the imaginary world (IW), therefore trying to visualize the response of the real world. Next this output is put through the intuitive section again and if satisfactory is outputed as a real action.

Or the information flow becomes:
W >input> I >feedback> IW >expectations> I >output> W

Last edited by absk @ 4/25/2016 2:29:00 AM

absk
posted 4/25/2016  02:27Send e-mail to userReply with quote
 
absk wrote @ 4/25/2016 2:24:00 AM:
Another thought is that for a very simple stimulus response model. Like the Deepmind Atari game playing programs. We only have two entities, namely the World (W) and the intuitively creature (I). Therefore the world is perceived by the creature and it immediately reacts.
However at Deepmind this however does not work well for the game PacMan since a sense of the world is missing, or in other words the walls are in the way...

A more complex creature would add a third component namely a imaginary model of the world (IW).

This more complex creature would still perceive the world. Come up with an intuitive response but next does not immediately output this response but gives this as an output to the imaginary world (IW), therefore trying to visualize the response of the real world. Next this output is put through the intuitive section again and if satisfactory is outputed as a real action.

Or the information flow becomes:
W >input> I >feedback> IW >expectations> I >output> W

 
Enter your message here

  1  
'Send Send email to user    Reply with quote Reply with quote    Edit message Edit message

Forums Home    Hal and other child machines    Alan and other chatbots    Language Mind and Consciousness  
Contact Us Terms of Use