I don't really think you should comment on something that you clearly don't know much about.
Try watching this for starters:
Human Rights For Robots.
- Thread starter tracertong
- Start date
I don't really think you should comment on something that you clearly don't know much about.
Try watching this for starters:
Has anyone created a version that is capable of recognising enemies and learning how to beat them?
For example the fire flower things, would it be able to understand that i needs to time its jumps to avoid the fire balls it spits, or to even wait before proceeding.
Scientists realized in the 60s that it wouldn't be possible to develop AI through explicit programming. You are only about 5 decades behind the times.
Silver being on the blob actually explains so much
for a game like that they'll generally optimize to become better than a human player
how it goes about it depends a lot on it's reward function* - does it want to complete the level in the fastest time for example or is it trying to collect all the coins/whatever
it is likely going to learn fairly quickly that it needs to avoid those fireball things but how it optimises for that might vary
*another poster mentioned emotion earlier but was unwilling or unable to expand on his comment - but you could look at the reward function as essentially the emotional drive/desire of the algo - it essentially wants to maximise it
Last edited:
About rewards.
Is there a system in existence which *meaningfully* defines its own reward?
Because that's something humans do. A trainspotter may get pleasure/reward from observing a diesel loco, and make efforts to seek out such trains. Someone else may have decided trainspotting is boring, and thus does not consider seeing a train to be any kind of reward.
So far in the examples I've read, the reward is pre-defined.
Also having the machine "randomly" choose a reward would be meaningless, and obviously possible. Ie, you could trivially define a set of rewards and have the machine pick one.
So how will we create a machine where the rewards are *not* pre-defined, and have the machine through experience, learning, environmental stimuli, decide for itself what is and isn't a reward?
Is there a system in existence which *meaningfully* defines its own reward?
Because that's something humans do. A trainspotter may get pleasure/reward from observing a diesel loco, and make efforts to seek out such trains. Someone else may have decided trainspotting is boring, and thus does not consider seeing a train to be any kind of reward.
So far in the examples I've read, the reward is pre-defined.
Also having the machine "randomly" choose a reward would be meaningless, and obviously possible. Ie, you could trivially define a set of rewards and have the machine pick one.
So how will we create a machine where the rewards are *not* pre-defined, and have the machine through experience, learning, environmental stimuli, decide for itself what is and isn't a reward?
it is called unsupervised learning.
furthermore, in Humans are largely preprogrammed in genes. Humans are rewarded for eating, frinking, having sex, keeping at a comfortable temperature, seeking shelter
Last edited:
Those basic drives are not what sets us apart from other animals, however.
There are many rewards which we develop purely via experience, observation, reasons unknown. Higher level pursuits such as artistic endeavour, etc.
Who knows why someone decides to take up landscape painting, for example - and why their brain starts rewarding them for doing so.
This kind of spontaneous change, with no direction or prompting, is something I wouldn't expect to see in a machine. Because we're not just talking about minor changes in a certain behaviour, we're talking major changes in personality, etc.
Or even something like giving up smoking. The choice to exercise willpower to stop responding to an established reward mechanism. We know that a smoker who quits has to fight his own brain's desire for more nicotine.
I can't imagine ever seeing such things occurring in a machine.
Maybe I've missed understood you or you have misunderstood me. But it seems to me that the program doesn't actually have an understanding of what the buttons its presses are doing. It seems more like its pure luck and a bit of brute force got it to the end of the level rather that intelligent progression/pathing through the level.
I understand that this is one guy with a little bit of time on his hands, but are there any examples of machine learning leading to intelligent progression to the reward.
i think you need to define "understanding"
understanding is merely the processing of information. Someone understands English due to electrical signals passing through neurons. A computer understands English due to electrical signals passing through transistors. A turing complete computer can conpute anything thatvis computable and there usno evidence that the human brain has super-turing properties.
actually thism raises a very intersting question.
I wonder what machines will think is beautiful.
oicne they reach a point iof having that concept.
i wonder what they will seek.
what is the difference between a human brain and an animals? Are animals conscious? If not, how did conciousness evolve in humans? if onlly humans are conscious then consciousness must have an incredibly simple neurophysiological basis since it evolved within a few million years compared to the billions it took to arrive at mammalian life in general.
There is nothing interesting about humans having hobbies. e.g, if a child was given a Thomas the Tank engine boxset then its no wonder they grew up with an interest in trains. Hunans arent thge onky species to paint, lots of artwork from chimpanzees, they spontaneously ask to paint.
Last edited:
I'm not sure what you mean, the algos will build up an 'understanding' of what happens when you press buttons fairly quickly, they also need to explore the environment, try different paths etc..
I know all I need to know, it takes purposeful intelligent design mechanisms from human beings to program complex man made software, even the blueprint comes from human intelligence to bring into existence complex machines, although some is copied from natures vast complexity.
but you made this statement:
which sigma picked up on - you seem to have missed that AI has moved rather a lot from logic, expert systems etc..
Well we'll certainly be asking them questions that we don't know the answers to!
I recently read The Last Question: http://multivax.com/last_question.html
Quite profound.