[LOCKSUIT]

Korrelan, your walk-through reply #2 is incomplete, long, and bad choice of words.

Your walk-through basically only says:
1) It generates actions for all motors.
2) Has a speed-o-meter reward.
Then what, magic? The hell this thing uses its actions.

This is why I sometimes just, I can't..

[keghn]

it come down to cause and effect.
A self learning bot starts with a single script program with address to its motors and the values it can pules these
motors pulse length of power. Say one pulse is a tenth of a second.
 Just check the motors you want to activate and then run the  program.
It randomly pulses a motor and then recorded the reaction, Once everything has settles down it make it make a copy
of the script program and then make mutations to the new copy. The old copy is for a back stepping.
It got to be in a tamed way one motor at a time and a low output value and pulse length.
The program that mutates it and auto runs the run scripts will not be in the run script it will be OFF the run script, like off
Off spark program.
putting the recording of events could be recorded in another place and a have a mark in the recording or where each
run script will activate.

 Everything is measured,. The wheels have encoders with optical pickup.   

https://www.google.com/search?q=image+encoder&espv=2&biw=1341&bih=675&tbm=isch&imgil=af0v7dParMU4DM%253A%253BemPpnQpNlcsNeM%253Bhttp%25253A%25252F%25252Fmechatronics.mech.northwestern.edu%25252Fdesign_ref%25252Fsensors%25252Fencoders.html&source=iu&pf=m&fir=af0v7dParMU4DM%253A%252CemPpnQpNlcsNeM%252C_&usg=__Mfh_HFOahIToqmj0qK7OskS3m_8%3D&ved=0ahUKEwj2tpv63IjPAhVBSGMKHaFECacQyjcIQA&ei=YwvWV7b5HcGQjQOhiaW4Cg#imgrc=TxOF3cNUZDVsrM%3A


 IN a un AI way,
you could just hard code it to do what you want. Like power the motors to do eight rotations then stop and turn right
and keep pulling information from your what ever collision detectors you have on the bot.

The raspberry pi has input output ports, GPIO, address tha have a program keep looking at and then make
the right move if that port goes high. like output to a turn motor.

[Freddy]

Korrelan, your walk-through reply #2 is incomplete, long, and bad choice of words.

Your walk-through basically only says:
1) It generates actions for all motors.
2) Has a speed-o-meter reward.
Then what, magic? The hell this thing uses its actions.

This is why I sometimes just, I can't..

Finally ! Now you know what we've all been going through with you these past few weeks   

[kei10]

Art, I never said you didn't know your aerodynamics nor that you can't fly planes, I was only using it as a example to say what I'm seeing from everyone in the AI field rather.

Yeah, we know. But Art that is just telling you that he flew a plane before you were born, and I don't see any problem with it.

Firstly, I forgot to say that korrelan explanation (reply #2) never even came close to fully explaining how it learns to move forth. See how mine did below.

Firstly, I should say that everyone has their own way of explaining things, and I do agree that providing a list is much more organized. The latter is that it depends on how well one has explained. I perhaps have no problem with your 8-steps.

By common sense, we're discussing about "how a spider-robot would learn to crawl on the floor". That says we don't necessary need to provide something like; "You place a robot on the floor", nor "Turn it on", because it is obvious that electronics needs electricity. So were that the spider's target is to move forward, and the responding variable is how fast it walks. Unecessary steps can be stripped off.

Additionally, a spider-robot's legs can be moved independantly with freedom. Unless its motors and legs are anchored and geared in such a way that it will always walk perfectly, then there's no point teaching it anything. Just shove in some batteries, power it on, and it will does its job.

Furthermore, I am not sure where you are going with this. The only difference between you and korrelan's explaination is that korrelan described a different algorithm than yours.

Please don't tell me you're still holding onto the "algorithms for all machine learning is the same!" belief ...

Plus the 20% he did explain, was so long, unclear, and isn't the best choice of words to use to explain it. Please notice how your explanation was NOT full.

I don't know what's up with you, but I don't see any problem with korrelan's words. There were no bad choices. And were very clear. I could instantly map a tree with ease.

Let us try to map korrelan's sentences into a list, shall we? And let's see if there's anything missing, incomplete, or so you claimed...

• The commonest method is to combine both the power of a neural net and genetic algorithm/ evolution
  
  • General information, yours are missing.
• The only pre-set required is target value (say speed forward)
  
  • "It starts generating random actions for all motors, and are saved as part1 of a single scene."
• The neural net is used to move the legs guided by a servo map, the servo map takes its operational parameters for a simulated G-Nome.
  
  • His are guided by neural networks(computation), servo map memory to servo-motor coordination (internal), iteration (supervising), and target speed. From his context, the servo-map is used to guide the servo-motors.
  • Yours are guided by "senses", "reward", and motor-coordination memory? Your steps are missing on what type of motors to be used, but nevermind that.
• A generic servo/ action map is generated along with forty nine or so (the more the better) copies with slight genetic differences.
  
  • The generic servo map is divided into fourty nine servo maps with slight variations, these variations perhaps based on estimations.
• Each map is loaded into the bots neural net and the servo map is run.
  
  • Rather putting on the floor and turning it on, this is more about the internal operation of the spider-robot.
  • "Yours would rank the tried actions, but when does it do them? When it sees cue! Cue is linked to tried actions."
• The two maps with the highest ‘fitness’ at reaching the target are then ‘bred’/ combined/ averaged and another forty nine copies are produced with slight genetic variations.
  
  • "Sensory input passes rewards and happens to trigger the accelometer reward. This scene of senses (part2) is linked to the actions tried, and ranks them. Now when sensory input enters again, it searches and selects the highest ranked but similar sense and initiates the linked actions"
  • Except that korrelan's method involves breeding two of the best servo map found.
  • The new-bred servo map is then made into new forty-nine copies with slight variations. That naturally means that the generic servo map is also inside the collections, a total of fifty. This is a rinse and repeat step... And you know where to begin.
• Rinse and repeat until the highest fitness measure obtained.
  
  • "The actions it just selected were tweaked, this continues improvement"

Yours would rank the tried actions, but when does it do them? When it sees cue! Cue is linked to tried actions.

The reason why you are not properly seeing that because korrelan's algorithm is completely different from yours. His algorithm involves generating copies with altered generic variations. The run for each servo-map is the record for the cue, the distance traveled, and the speed record. Two best copy is the required cue, which it is then combined, and the very steps is repeated.

If you are truly interested in this, you shouldn't have any problems with it, especially appreciating answers from others.

I feel like I've been giving advises non-stop... Please learn to read before throwing the blame. 

[LOCKSUIT]

1. The commonest method is to combine both the power of a neural net and genetic algorithm/ evolution
General information, yours are missing.

My algorithm does have this General Information - the way it works is described and the neural net is not needed since any functions it does can be put into a metal box with wires lol. As for genetic algorithm, mine is not grown from DNA, rather our hands and hammers.

"rest"

Well, ok, it does the job of learning to crawl, BUT, do you realize korrelan's robot simply keeps doing the crawl actions after tweaking their power and directions numbers? It's ONLY supposed to select a motor map of actions for ex. hand ex. hammering/pointing/stirring/drawing on the cue, not simply initiate n tweak the same one allll the time.

But still yes, I explained the same thing (but more advanced) with better wording, and shorter text. Motor map of all motor #s are in my text, don't worry, the tweaking repeat is all their. Korelan's/the field's just reads like a caveman's Egyptian encryption explanation that's not fully understood. This thread was to show that, it worked in a way...at the beginning I expected the answer to be similar to my 8step one.

Admit it, the field calls "assorting senses" "labeling data". The brain is saving images and are being saved near look-alikes for assorting (for fast searching (by directing input to it)). A 3D/2D assortment of images by color, brightness&shape is NOT "labeling" it is soo "assorting" plus is why the cortex is calllled Association/Associative tooo, Primary is primary colors/brightness tada. Lastly we know images are data, but that word mustn't at-all be used, only use "images", = assorting images = makes total sense. Labeling data.

[kei10]

This thread is about "Explain to me how a spider-robot would learn to crawl on the floor.".

But yet you seems to have been directing the discussion into "Who's algorithm for how a spider-robot works best" after all this time ever since the others have already answered what you wished for.

Admit it, for the worse part, you're continuously preaching that yours works best (in every single related thread). All algorithms has their own use, big-o-notation, different performance, and efficiency level.

Please learn to grasp the situation. 

[keghn]

Yes simple robot spider. But i can see a spider that  keep learning all the way up to AGI and beyond to ASI, artificial
super intelligence. 

I see a path to do it. 

 Most do not want their creation have a primitive beginning. Very un glamorous for a super AI to find out it came from
nothing and not  a clone of royalty. Not puffed  into existence by a god. But by a race that  needs to model, to work out their
imperfections.

 Most want to hard code robot with every action it. Or take a IBM Watson clone and convert it into a AGI. Very
glamorous, safe move.

[LOCKSUIT]

Yes keghn you can't create a super AI instantly, it has to learn, in a very language-able body not inside of a computer. The body has to be right, the algorithm, and what it learns.

Kei/all, your taking the thread name exactly, again, it is about if you really "know" machine learning in a better way....that's why I gave a example above about how all AI field people say "labeling data" when it should be called assorting senses as explained why. If you stick to calling the terms names that refer to other meanings like data when it is images you are dealing with them you are onnllly focusing on the wrong thing, get rid of data word, only use images/senses. There's sensory cortices etc, and the algorithm you describe can use the word senses/reward - it works for any algorithm, though they all will do the same thing ex. learn actions & cue etc, trying to tell yous. Again, don't focus on neural nets, rather metal boxes with cords and the functions you need inside.

[kei10]

And certainly different terminologies are very important. If you can't agree, then I'll just agree to disagree. Fine. Have it your way.

However, "Do You Know Your Machine Learning" does not meant "Who's algorithm for how a spider-robot works best".

[spydaz]

What a mad thread !

An ai can't learn anything that has not been previously completed.

As it does not know what crawling is and needs a comparison to the final outcome ...

With the final outcome known it's just a matter of how long will it take to reach the outcome desired ....

There are multiple ways of completing the task ; but as previously mentioned the simplest is a neural network ; a recurrent neural net ... As its a sequence that is being predicted .... A Markov chain could also be used ...
As it too remembers sequences . 

Machine learning is not a solve all ....

This problem is a firstly supervised learning problem . Then after being solved by the algorithm then it can be applied to "different surfaces" to see how quick it could learn to walk / crawl on that surface ... This is the idea behind this problem ...
In reality if you was designing a "dolly" which could crawl you would pre program it and not use machine learning ..

Just as with the cars which park themselves. It know how to park ... But it needs to scale that parking to any space that fits .... This is where the learning is required ... Simply train on the neural network with different measurements until finally any measurement which fits it will successfully park .... Knowing the error margin allows for knowing the minimum safe parking space which will yeild 100 percent success and refuse all other spaces even if it can fit . Then it becomes a unsupervised algorithm ....

It's always possible to use reinforcement learning , but why ? It's the wrong way.

Although an algorithm was proposed as the answer ... Obviously it's coming from guesswork and expecting kudos for the wrong answer .....

Also insulting members because their answer does not fit your university criteria (obviously come on here to get answers for your school project) is bad etiquette ... Perhaps pose your uni questions on stack overflow... Or search git hub ...

Actually the answer is on the MIT website ....and there are many papers explaining exactly what I said although long winded !
Research and read !

You should also know when designing a robot you would use the intelligent agent paradigm ... That's book 1 in the artificial intelligence business .

Stuart Russel and Peter norvik
Google it!



Sent from my iPhone using Tapatalk