Use my link bit.ly/NeatAIDCJan22 and check out the first chapter of any DataCamp course for FREE to unlock new career opportunities and become data fluent today!
I love novelty search. I'd argue, though, that "novel" is third to "interesting". - Second would be "Surprising" - which combines being novel with being difficult to predict. And algorithms for that do exist. For instance, you could ask a neural network to predict what happens next, and then prefer the things that the network was most off about. It's still trickier than novelty though. You gotta worry about the stability of such a network, and not all things are easy to predict. But for the maze example, one method would be to predict a heat map of the next generation, and and then estimate how likely the actual next generation's samples were drawn from the same distribution, ranking them by how likely they are to be outliers.
Cool thing to do would be to run the Pic breeder images through an image recognition NN, and just selecting the image with the highest certainty rating for anything at all. Maybe you will end up with several images classified as the same thing, then use the top few of them instead of the top one. Could be a way of searching for something, but not in particular.
We really need channels like this, with compressed overview of science papers on interesting topics, or kinda inspiring ones like Novelty search. It's interesting as hell.
I wonder if it were possible to expand on picbreader by introducing encapsulation Instead of just a handful of starting functions, it could be interesting to also allow for already found functions from different branches to be applied to each other. Maybe using some surprisal search type deal to steer what particular combinations are gonna be shown
Maybe combine novelty with fitness by using the inverse fitness to define the novelty requirement. So if a solution is miles away, the novelty radius is large, but “nearly there” gets finer search.
I disagree with the novelty search idea - it seems to me that it can only work in a very low dimensional world... If the number of potential novel solutions is infinite, no progress will be made and keeping track of what has been visited becomes prohibitively expensive. I've always argued (even had this discussion with the author of the paper) that it seems like a weak substitution for a more finely grained fitness function. Meaning if the fitness was instead how close you got to the solution was part of the reward, that would also make the problem easier to solve. I may be wrong though, since lots of papers were written on the subject. As a massive sidenote that will only entertain me - I'm actually mentioned in the Picbreeder paper. I was also the first person to evolve images using NEAT, but what I was doing was copying others who had been using Genetic Programming to do the same. My implementation has no relation to their implementation, so the credit is all theirs.
For the milestone approach, are the resulting agents able to navigate the maze from start to finish? The approach / examples shown in the video lead me to believe that they would be just as lost at the start of the maze as the standard approach.
yes, they'll do it from start to finish as they're rewarded for finding something new.. which causes them to disperse throughout the maze.. But stepping stones is equally valid in this case.
I think the green balls should have been shaped like cat heads.. since it seems how cats brains are wired. You rearrange furniture or open a door to an area they don't normally go to, and they are compelled to explore this "novel" space.
Does this scale to higher dimensions though? I have the suspicion this would start providing less and less advantage the more dimensions the maze has...
Hi, it will be a bit unrelated and weird but i was wondering; could a time travel-like structure be made with cellular automata? Like, cells affecting the ones from previous generations? What are your thoughts?
Click bait titles for this could be: "AI demonstrates the importance of diversity for its own sake" or Daoist AI shows that the only way to find your goal is not to search for it."
Use my link bit.ly/NeatAIDCJan22 and check out the first chapter of any DataCamp course for FREE to unlock new career opportunities and become data fluent today!
That segue into the sponsorship was smooth 😂
I love novelty search.
I'd argue, though, that "novel" is third to "interesting". - Second would be "Surprising" - which combines being novel with being difficult to predict. And algorithms for that do exist.
For instance, you could ask a neural network to predict what happens next, and then prefer the things that the network was most off about.
It's still trickier than novelty though. You gotta worry about the stability of such a network, and not all things are easy to predict.
But for the maze example, one method would be to predict a heat map of the next generation, and and then estimate how likely the actual next generation's samples were drawn from the same distribution, ranking them by how likely they are to be outliers.
The first video about science that gave me actual advice for living a better live. Thank you :)!
Glad it was helpful!
Cool thing to do would be to run the Pic breeder images through an image recognition NN, and just selecting the image with the highest certainty rating for anything at all. Maybe you will end up with several images classified as the same thing, then use the top few of them instead of the top one.
Could be a way of searching for something, but not in particular.
I thought they tried that with the eye and it took 29,000 steps to get somewhere that looked worse than the original 12 step eye
@@RobinDuckett i think the issue was trying to make it look like an eye specifically, instead of anything
We really need channels like this, with compressed overview of science papers on interesting topics, or kinda inspiring ones like Novelty search. It's interesting as hell.
I appreciate the intriguing information, but also the philosophy too!
This is amazing!
I love your channel! Thank you for all the effort you put in
Thanks for watching!
so all things really do exist within reality, its just up to us to discover them
With this method we'll crack physics and mathematics... I hope.
I wonder if it were possible to expand on picbreader by introducing encapsulation
Instead of just a handful of starting functions, it could be interesting to also allow for already found functions from different branches to be applied to each other. Maybe using some surprisal search type deal to steer what particular combinations are gonna be shown
Maybe combine novelty with fitness by using the inverse fitness to define the novelty requirement.
So if a solution is miles away, the novelty radius is large, but “nearly there” gets finer search.
good idea !
I think they way they learned to levitate without using energy is where you buried the lead.
Just found your channel and already absolutely love it, keep up the great work
I disagree with the novelty search idea - it seems to me that it can only work in a very low dimensional world... If the number of potential novel solutions is infinite, no progress will be made and keeping track of what has been visited becomes prohibitively expensive. I've always argued (even had this discussion with the author of the paper) that it seems like a weak substitution for a more finely grained fitness function. Meaning if the fitness was instead how close you got to the solution was part of the reward, that would also make the problem easier to solve. I may be wrong though, since lots of papers were written on the subject.
As a massive sidenote that will only entertain me - I'm actually mentioned in the Picbreeder paper. I was also the first person to evolve images using NEAT, but what I was doing was copying others who had been using Genetic Programming to do the same. My implementation has no relation to their implementation, so the credit is all theirs.
For the milestone approach, are the resulting agents able to navigate the maze from start to finish? The approach / examples shown in the video lead me to believe that they would be just as lost at the start of the maze as the standard approach.
yes, they'll do it from start to finish as they're rewarded for finding something new.. which causes them to disperse throughout the maze.. But stepping stones is equally valid in this case.
I think the green balls should have been shaped like cat heads.. since it seems how cats brains are wired. You rearrange furniture or open a door to an area they don't normally go to, and they are compelled to explore this "novel" space.
It's an enemy stand!
Wait that website is down? I had used it and already forgotten about it
Does this scale to higher dimensions though? I have the suspicion this would start providing less and less advantage the more dimensions the maze has...
Duuude i remember how i wanted to use picbreeder for something graphic design related, but i couldn't find it
yea, picbreeder was very nice.. easy enough to code up for home use..
Hi, it will be a bit unrelated and weird but i was wondering; could a time travel-like structure be made with cellular automata? Like, cells affecting the ones from previous generations? What are your thoughts?
as part of an evolutionary algorithm? Or IRL?
the blind watch maker --
Click bait titles for this could be: "AI demonstrates the importance of diversity for its own sake" or Daoist AI shows that the only way to find your goal is not to search for it."
Nice.. will have to give them a go..