If you want to make a program that makes choices like a living thing.
-
wrote 17 days ago last edited by
If you want to make a program that makes choices like a living thing. Model it after a living thing.
Obviously, modeling a human mind would be hard, but what about an ant doing one task? Lots of people have done this, mostly focusing on large numbers of automata following very simple rules. Dead simple rules. Too simple.
Ants are pretty complex. For one thing they have "emotional states" or if the word "emotion" bothers you, you can call it "a persistent condition that influences choices." 1/
-
If you want to make a program that makes choices like a living thing. Model it after a living thing.
Obviously, modeling a human mind would be hard, but what about an ant doing one task? Lots of people have done this, mostly focusing on large numbers of automata following very simple rules. Dead simple rules. Too simple.
Ants are pretty complex. For one thing they have "emotional states" or if the word "emotion" bothers you, you can call it "a persistent condition that influences choices." 1/
wrote 17 days ago last edited byAnts, at minimum, can be "alarmed" or "scared" to one degree or another. This changes how fast they move and how likely they are to attack. Ants can be more or less cautious and this is correlated to the size of their colony. The bigger the colony the more bold the ant.
Just having ants seek food and lay trails after finding food isn't enough to get the kind of complex emergent behaviors you see from real colonies.
Give your digital ants "emotions."
2/
-
Ants, at minimum, can be "alarmed" or "scared" to one degree or another. This changes how fast they move and how likely they are to attack. Ants can be more or less cautious and this is correlated to the size of their colony. The bigger the colony the more bold the ant.
Just having ants seek food and lay trails after finding food isn't enough to get the kind of complex emergent behaviors you see from real colonies.
Give your digital ants "emotions."
2/
wrote 17 days ago last edited byBut that is not all. They also need memories and the ability to learn by association.
And lastly, they need persistent, but changeable goals.
We seem to be delighting in wasting processing power, I see no reason not to really dig in and try to model some ants and see what emerges.
3/3
-
But that is not all. They also need memories and the ability to learn by association.
And lastly, they need persistent, but changeable goals.
We seem to be delighting in wasting processing power, I see no reason not to really dig in and try to model some ants and see what emerges.
3/3
wrote 17 days ago last edited by@futurebird I'm pretty sure I sent you this before, but in case I didn't:
It's a real thing ... pheromones, trails, etc.
-
@futurebird I'm pretty sure I sent you this before, but in case I didn't:
It's a real thing ... pheromones, trails, etc.
wrote 17 days ago last edited byYeah I've read this book. It's more about trying to isolate the few simple rules that produce solutions in systems with multiple agents. And avoiding making those agents individually "needlessly" complex. It's what made me wonder what if you didn't focus on that and tried to make the ant agents more ... like real ants.
-
Yeah I've read this book. It's more about trying to isolate the few simple rules that produce solutions in systems with multiple agents. And avoiding making those agents individually "needlessly" complex. It's what made me wonder what if you didn't focus on that and tried to make the ant agents more ... like real ants.
wrote 17 days ago last edited by@futurebird As far as I know the people working in this area - meta-heuristics for complex optimization - mostly use methods derived with computer science / discrete math, rather than by modeling living organisms or colonies of them. They have test sets and competitions just like the AI folks do.
-
F myrmepropagandist shared this topic 17 days ago
-
@futurebird As far as I know the people working in this area - meta-heuristics for complex optimization - mostly use methods derived with computer science / discrete math, rather than by modeling living organisms or colonies of them. They have test sets and competitions just like the AI folks do.
wrote 17 days ago last edited byIt is perhaps pragmatic to boil down the few essential rules that help ants be efficient.
But, I've always been curious about "life like" models. Because all of these creatures we kind of hope are "simple" ... aren't.
-
It is perhaps pragmatic to boil down the few essential rules that help ants be efficient.
But, I've always been curious about "life like" models. Because all of these creatures we kind of hope are "simple" ... aren't.
wrote 17 days ago last edited by@futurebird How many neurons does a worker ant's nervous system have?
-
@futurebird How many neurons does a worker ant's nervous system have?
wrote 17 days ago last edited byThey have rather large brains among insects given their body size, although bees beat them by a little. They have from 100,000 to 350,000 neurons. Ant size and the size of the eyes are a big factor.
-
They have rather large brains among insects given their body size, although bees beat them by a little. They have from 100,000 to 350,000 neurons. Ant size and the size of the eyes are a big factor.
wrote 17 days ago last edited by@futurebird So 350K neurons times the number of ants in the foraging group - that's in the millions of neurons, I think. That does seem more efficient than a meta-heuristic.
-
But that is not all. They also need memories and the ability to learn by association.
And lastly, they need persistent, but changeable goals.
We seem to be delighting in wasting processing power, I see no reason not to really dig in and try to model some ants and see what emerges.
3/3
wrote 17 days ago last edited by@futurebird I love this as a metric... Especially around the AGI folks.
"Oh you're working on an AGI? Can you reproduce an ant colony?... What, you can't model an ant colony, but you're confident that you've made a human intelligence? Can you even model an ant?"
I especially like this bar because it's easy for most people to grasp. If the dude in front of you says they have an AGI, but they can't model an ant, then you know they're lying.
-
@futurebird So 350K neurons times the number of ants in the foraging group - that's in the millions of neurons, I think. That does seem more efficient than a meta-heuristic.
wrote 17 days ago last edited bySome colonies have... millions of ants.
-
@futurebird How many neurons does a worker ant's nervous system have?
wrote 17 days ago last edited by@AlgoCompSynth @futurebird The connectome is a very small piece of the puzzle that would be required for emulating any organism's brain in any significant complexity.
We're only now beginning to simulate some of the behaviors of C. elegans using realistic models (OpenWorm and BAAIWorm for example), and that's a far less complex organism than an ant.
For example, we have several Drosophila connectome datasets at this point (released after the linked paper was published), but we're still far from simulating D. melanogaster. This paper is a good explanation of roughly what we still need before we could begin a OpenWorm-esque project for D. melanogaster.
Depending on which paper you read, ant species tend to have slightly to significantly more complex connectomes than Drosophila.
