Episode 13 - Game of Drones

In this episode of Exploring AI Matters we discuss swarm behavior in social organisms like bees, termites and ants.  These emergent behaviors point the way toward possible military applications for swarms of drones or other robots.

Our guest for this episode is Lieutenant Colonel Jason Cody from the United States Military Academy at West Point.  Colonel Cody’s doctoral research at Vanderbilt focused on swarm intelligence and its potential in practical applications in human-swarm teams. [2023-08-15]

Transcript

SPEAKER_00 0:03

Welcome to Exploring AI Matters. This podcast series, previously known as Mind the Gap, Dialogues on Artificial Intelligence, will continue to appear in the ABA series to the extent that in addition, all of the episodes, old and new, will now appear under our new podcast name, Exploring AI Matters. Thank you.

SPEAKER_01 0:37

Humans tend to think of intelligent behavior as originating in the brain. Analogously, we think of artificially intelligent behavior as originating in a computer central processing unit. In today's episode of Mind the Gap Dialogues on Artificial Intelligence, we shall discuss intelligent behavior emerging from swarms of simple robots. In nature, we see coherent cooperative behavior from collections of animals like colonies of termites, colonies of ants, and hives of bees. Scientists have studied these organisms and their behavior for a long time, trying to understand how these animals coordinate their behavior so effectively, given the simplicity of their nervous system. Computer scientists began to explore swarm intelligence in the 1990s. Our guest for this episode is Lieutenant Colonel Jason Cody from the United States Military Academy at West Point. Colonel Cody's doctoral research at Vanderbilt focused on swarm intelligence and its potential and practical applications in human swarm teams. Jason is an assistant professor and active duty officer, teaching in the Academy's Department of Electrical Engineering and Computer Science. As a career signal corps or communications officer, he has deployed to Afghanistan, Iraq, and Kuwait. In addition to his distinguished service in the Army, he has authored and co-authored numerous academic publications in artificial intelligence and is a graduate of the Command and General Staff College. Hello, I'm Roland Trope, a national security lawyer.

SPEAKER_04 2:21

And I'm Charles Palmer, a computer scientist. We are your hosts for this episode of Mind the Gap Dialogues on Artificial Intelligence. In addition, we have two more hosts.

SPEAKER_00 2:33

Hi, I'm Ama Adams, a national security lawyer.

SPEAKER_03 2:37

And I'm Mark Donner, a computer scientist.

SPEAKER_01 2:40

Each episode will be led by two of us, with the other two adding impromptu questions and comments as the spirit moves them.

SPEAKER_04 2:48

So, Lieutenant Colonel Cody, thank you for your time today. It's good to see you again.

SPEAKER_06 2:52

Before I go in any further, though, I do want to throw a disclaimer. So, of course, I am uh you know assigned to the United States Military Academy at West Point, uh, but I do have to clarify that uh I'm here in this interview expressing my own personal views and my personal capacity. Uh any of any of the opinions or views or conclusions that I share today are my own and are not to be interpreted as representing official policies or endorsements of the United States Military Academy, the United States Army, the Department of Defense, or the United States government. Sure.

SPEAKER_04 3:22

So help us understand this uh intelligent coordinated behavior in swarms of animals.

SPEAKER_06 3:29

Um very excited about it. All the uh all the uh animals out there coordinate in some way. Um, you know, humans coordinate, of course. You can think of the higher-level organisms uh commute uh uh coordinating. What we're talking about here when we talk about swarms, we talk about large groups of typically relatively what we consider simple animals, uh, although they can be quite complex in their own way, coordinating together to solve some important task. Uh and so a couple of examples I think may help. Um, one is uh you know, fish schooling, for example. You know, today we're gonna talk a lot about social insect colonies, but fish schooling, uh, you can imagine uh uh schools of anchovies in a swirling mass that looks almost like a liquid living thing all by itself, even though we know it's made up of thousands of uh individual organisms. And uh this this superorganism, if you think of it that way, is able to respond intelligently to its environment and find food and avoid predators, all because the individual fish are able to uh to to exploit the fact that a fish within the organization can see the predator even if I can't, for example. And so that that swarming mass occurs because in each individual fish is saying, okay, I'm going to avoid the neighbors next to me. I'm going to uh align myself with neighbors that are further out, and then I'm going to try to move towards those that are further away. And those fairly simple movements are all that's required to create these really complex behaviors that have then been replicated by computer scientists and some simple simulations. The real fish are more complex than that. It is still interesting to see that simple rules result in these emergent behaviors that if you were to look at any individual fish or even a small group of fish, you couldn't really see what the whole effect was. And uh, nor does it make sense to say that an individual fish is avoiding the predator if the predator is nowhere near that individual fish, if that makes sense. Sure.

SPEAKER_04 5:23

So, okay, I I kind of get the idea for the predator issue, but does it help them in any other ways? I mean, termites, ants, bees, uh, you know, I'm not sure they have the same kind of predatory issues that the fish do.

SPEAKER_06 5:37

Yeah, that's a different type of behavior. Um, and in in some ways it's much more complex. So the fish, you know, they're largely out for themselves. It's kind of selfish. They find food because fish that know about food uh go in a certain direction and get their neighbors to follow them and ultimately move the whole mass. Whereas a colony of termites or ants or bees is really engaged in an interaction with the environment and with predators, but also with food sources and maintaining their own systems uh at a con at a constant period of time. So their behaviors, you could decompose them. You can think of the individual colonies as getting their own food, regulating their own environments, whether that be through construction activities or through uh through removing waste from the colony, expanding the colony, moving and taking care of the babies, the nurseries, and of course taking care of the queen, which does not control everything. Uh, and so you have several different behaviors that each one, each single behavior in its own is a separate swarming behavior. But this complex interaction between these behaviors is interesting because it allows them to survive as a really a superorganism, um, which would not be possible for, say, a single ant on its own.

SPEAKER_04 6:48

Well, that that single ant makes me ask another question before we continue.

SPEAKER_06 6:53

Sure.

SPEAKER_04 6:54

In human organizations, one uh soldier asleep at the wheel or one driver not paying attention can screw things up for the whole uh group. Do individuals in these swarms really matter if they kind of stray or if they're uh incapacitated?

SPEAKER_06 7:11

Now, one of the one of the fascinating things about these about insect uh organizations is that you know the the colony, with the exception of the queen, although the queen is not in charge, all the rest of the organisms are basically expendable to a certain degree. You can take many of the individual ants, regardless of what their task is that they're performing. And then the individual task and the colony as a whole is generally okay until they reach uh until you know it the colony was being attacked by another colony, for example, um, and it reached a certain critical level of uh you know a minimum population where it could no longer perform the storming task, then the task itself would fall apart and then the colony would be in danger. But that's a long way to go to take out the colony.

SPEAKER_04 7:53

Yeah, they okay. Well, let's let's do a little math here. So um I think we've said or we were told that ants have about a quarter of a million neurons, and bees, you know, maybe half a million to a million. And neurons are, however, kind of slow. I mean, ours are too. You know, taking about a millisecond to switch. How can they have such a sophisticated behavior with such slow and limited computational hardware?

SPEAKER_06 8:23

They they have the same limitations and overcome them the same way that we do. Um, so I I'd like to say, you know, I didn't study individual insect brains, but I am, of course, aware of their extraordinary capabilities. So the bottom line is that they have a really fascinating ability, just like we do, of interpreting overlapping sensory information and the interconnections between their neurons uh far outnumber the number of actual neurons in their brain. And so I don't have the number on me right now as to what those interconnections look like, but in a human brain, it's billions of cells to trillions of interconnections. And uh, and I assume uh from observing some of these that the uh the interconnections between the bees and the hierarchy uh with bees and ant brains is similarly organized, uh, like the human brain is, in terms of there are structures that control other structures of neurons. And that kind of hierarchy and that kind of interconnection uh means that even though each individual connection is slower, they're getting they're processing more in parallel. Um, and so good examples of uh, you know, bees are clearly able to, individual bees are clearly able to learn behaviors on their own. So, you know, one interesting study is, you know, we're gonna talk a lot about bees, I'm pretty sure, because that's what my what my uh research focus really was on was uh some honeybee behavior. But a worker only lives about three to five weeks, and yet this single, this single bee, uh, as it gets a little bit older in their terms, um, as it gets a little bit older, they begin doing more complex behaviors. And so a bee will remember how to get to a distant foraging site. Uh, they'll remember going to multiple foraging sites. Um, and that's a lot to think about. You know, this brand this bee is flying to a distant location. Uh, you know, they can go like that. I am I'm gonna butcher this number here, but I know it's you know, you can go kilometers, is what we're talking about, but they can go pretty far away and still come back without um without much of an issue. So they are doing their own land navigation and figuring out, you know, how and remembering where to go and where to get back, uh, not to mention the fact that they also know how to transmit that information to other bees so they can find it. Uh that's a lot of complexity for an individual uh bee, which is why it makes it so interesting to talk about uh what they can actually do. Yes, that's Jason.

SPEAKER_01 10:34

From your research, especially the focus on bees, uh what were you able to draw from your study of bees in swarms for use in AI? And I have to say, when I was reading through parts of your dissertation, I was really fascinated by your comments about how bees select as a swarm a new nest site. Could you explain some of that for us?

SPEAKER_06 10:55

Um that's definitely honeybees are are quite simply amazing. So uh what happened, what happened to me is I uh I picked up Tom Steely's book, The Uh The Honey Bee Democracy, which anyone is interested in bees, and hopefully I excite some of your listeners uh about bees in this next uh few minutes here. Tom Seely spent a large portion of his life really documenting uh how bees thrive and how they how the hive works. But most importantly in this book, he really looks at how the bees um make decisions. And there are a lot of different decisions to go into, but the behaviors that I've pulled from bees, probably the most important one, their ability to choose a new hive site. So if anyone's not familiar with this, what happens is you have a hive of you know the tens of thousands of bees uh in it, and there's a there's a queen, and the at a certain point, uh typically when a new a new queen is about to emerge, uh the old queen will, the hive will emit a swarm of bees. And that swarm of bees' job is really it's if you look at it the hive as a as a single organism, this is really the reproduction of the hive. It's gonna go out and form a new hive somewhere else. And so that is a well-known behavior, but what is not well known, and what was uh revealed in in Tom Sealy and his colleagues' work is exactly how that happens. So I'll focus on that behavior. Happy to discuss other behaviors too. But the uh the bottom line is once the uh once the swarm is emitted from the hive, it sets up kind of a temporary location. Um, this temporary location is uh can be hundreds of yards away from the original hive, and they form this mass that some of you may have seen, just clumped anywhere that makes sense, usually like in a tree or somewhere, uh, but it could be anywhere in a house, uh, wherever. And this this mass really exists to protect the queen. So it regulates temperature, keeps the queen protected. But at the same time, two to five percent or so of the swarm is made up of these little of these scouting honeybees. And these scouts will go, and their job is to look, look for and discover sites that are possible hive locations, uh, future hive locations. And these scouts will independently go and assess uh assess sites that they find. And these sites uh really matter in terms of the size of the opening. If it's too big, it's dangerous because it let out too much heat. If it's uh, I think the optimal size is about 40, you know, 40 liters of volume within the hive. Uh so they assess several aspects of these possible sites. And then the interesting thing is the scout that discovers it then goes back and has you know 500 or so of its uh other scouts that has to convince um of which site is best. So there's no centralized reporting procedure. So they actually start initiating this uh voting dynamic, really a recruiting dynamic. So an individual scout that sees a site will come back and it will, if it likes the site and it and it assesses it as being a good possible hive site, it will attempt to recruit other bees to go and assess that site. And so they have a dance floor, similar to you see uh the dance floor on the hive, where they uh equally compare foraging sites when they're not in a swarm. On the swarm, when they're comparing future nest sites, uh there's a location on a swarm where they they do this, perform this dance. And I assume I've got time to talk about the waggle dance because it is quite fascinating. So some of you may have heard of this. A honeybee that is trying to communicate information about a location will land on this uh this decision-making hub or this dance floor, if you will, and it will first align vertically with uh with the swarm and then turn so that uh so at a at a particular degree away from vertical. And then it will vibrate its body and move in a circular motion. Well, it is uh what is fascinating about that dance is that the turn it makes from vertical is really the change in azimuth, the direction from where the sun is with relation to the swarm, and then the number of rotations it makes is uh scalable to the distance to the site location. So you have an azimuth and a distance that's communicating to other bees that are watching this, interpreting it, and knowing where to go, which to me is fascinating. And Tom Scaly in his book talks about he didn't know where the skeletons were going and had no idea. And he was on this island and the bees weren't going where he expected them to go, and he watched them dance, and he could actually plot on a map where they were going, which is fascinating to me. So, this recruiting behavior, uh, the number of dances, how well they recruit, is actually proportional to how well the individual will be assessed that site. So better sites result in better dancing, which results in inspiring other bees to go and check out the same site. The other bees will go check out the same site if they if they agree with an independent assessment, they come back and also recruit. Uh, and so you can imagine this recruiting process alone is enough to start building support and varied degrees for different sites. In addition to this recruiting behavior, each individual bee also inhibits bees that it sees that are advertising other sites. Uh, and that inhibition is also proportional to the quality of the site that they're currently supporting. So if I'm B A supporting, uh supporting a site, and you're B you're a second B, uh B B and you're doing another site, I the probability of me inhibiting you from advertising your site is actually proportional to how well I assess my own site. And then this process continues until the uh each one of the bees, as they recruit, they they go back and forth between recruiting and then going back and reassessing their site. And that reassessment is fascinating because what they're really doing is not only improving their assessment of the site again, but they're also really counting the number of bees that are at that location. And when the number of bees is roughly between 20 or 30, then they've evolved this ability that, okay, that means about half of the total bees are assessing my site. And so we've got a majority, and they go back and start preparing for movement. Very complex answer. It is equally interesting problem to figure out how those bees then motivate the swarm to actually prepare and then take off and then move, you know, thousands of thousands of meters, a couple kilometers away, and then arrive at a very tiny spot altogether, even though only you know less than a half percent of them know where they're going.

SPEAKER_01 17:15

So what that answers about the first half of my question, which is what were you focusing on? But what did you find that was useful in what you've just described for application in AI? And I have some follow-ups to that, but I want to make sure you complete the bridge that you started building.

SPEAKER_06 17:31

What I was most interested in was the decision-making process itself. You know, how can a decentralized group of agents, whatever the agents were, uh limited to uh limited to local communication, how could they, in a self-organized manner, assess a discrete set of options uh and then choose between them, and then such that every one of the agents in the group would then ultimately be able to execute the same action. So in my research, I was uh I was certainly inspired by biology, working in with artificial agents, and these were simulated agents, by the way. Uh, and they were doing they were doing site selection assessments, and uh and then the whole group would would lift off and and then travel. But the idea being uh twofold. One is um I was exploring the decision-making process itself to see if it could be uh improved upon. Uh, to be honest, I think that the the bees have it pretty pretty good. The models we have so far are not perfect in terms of they're not as resilient as the bees themselves. Um, but also also looking at not just choices within, but also could uh you know, could these agents make a decision at all? So this is a best of end decision. In this case, if I'm assessing sites, that's end sites. If I'm assessing courses of action, that's n courses of action. And so if they can come to a consensus, that means that they have uh they have a decision-making capability in which not only does are they able to assess options, they're able to choose one. And then every agent in the group knows the knows what the outcome was, and then can from that point take action, if that makes sense.

SPEAKER_01 19:11

So it did, except I'm still not sure that I understand how do you make use of that in an AI program to accomplish a particular task.

SPEAKER_06 19:21

So let's see. So you're just saying, what is the uh how do I, I guess I'm not sure exactly what your how your question is um is asked then. So the AI, I'll I'll start with this to see if it satisfies um your question. If not, we can go into um a dive about some of the um some of the you know how you would apply AI to this process to optimize it. So for example, the early swarm uh experiments focused on capturing the behaviors at the individual level and then showing how those individual microscopic behaviors would cause emergent intelligent behaviors as a collective. So in this case, the individual, the individual agents that I was dealing with, those are very a fairly simple, more or less simple model uh for the assessment and then the recruiting and then the inhibition of other options. Whereas uh what it had to do is nest within a collective model that mathematically converged to an actual decision. And so these are those are mathematical models, and it's it's analogous to you know early forays into AI. And the next step for these really is okay, those interactions, those models you fall prey to uh exploring an exhaustive search of parameters, for example, which is very hard to control. You know, the the you know how many interactions per the quality of the site is uh is very difficult to cleanly and hand craft. So a lot of the a lot of the most recent um articles that have come out about storm intelligence, not surprisingly, are how do we go about feeding the right feature space into an artificial intelligent algorithm to define and learn that parameter space uh instead and improve the intelligent agents, which is a hard problem and currently unsolved. Um, but that is the direction I see this going.

SPEAKER_03 21:10

I think so so let me just uh interject uh a question I think suggested by Roland's uh sort of train of of questioning and your and your descriptions of what's going on. How sophisticated is the uh the decision-making behavior by the individual B that produces this emergent uh rather uh rather very sophisticated decision-making behavior? I think that's sort of the thrust of your comments.

SPEAKER_06 21:42

So I think for obviously the models that I explored were quite simplified. So they were doing uh, you know, relatively abstract evaluation of sites. Uh, if you mean the bee itself, the the bee's assessment of a site is actually quite complex. So, did you mean the biological assessment or the uh The simulation simulated videos.

SPEAKER_03 22:01

Yeah, I think so. In some sense, the complexity of what you're describing, you know, the I mean again, the B is a relatively simple engine. Uh, it doesn't, you know, uh the the the the the emergence of sophisticated behavior out of relatively simple things is what I sort of uh what I think you're getting at, and what I think Roland is trying to grasp from his questions.

SPEAKER_06 22:28

So I think um I'm still not sure if I'm getting there, and I apologize, but I think that the in a lot of these behaviors, and this one in particular, often often defining the individual complexity um is actually almost a separate problem to defining the complex problem that the uh the collective is trying to solve. Uh and so in this case, I could track a single agent or be you know throughout the entire process. And their behaviors at any given time uh don't appear that, you know, they're learned behaviors, um, but the collective as a whole can react to wildly dynamic environments as a result of these behaviors. So they can make comparisons between um between any number of, I wouldn't say any number of sites, because there actually is an ill-defined uh upper limit on the number they could actually um assess. But they can also respond dynamically if a site is destroyed mid-decision-making process, or um or even if one is one is moved, it can be rediscovered. And that is that is quite complex. Um, and the collective is able to accomplish that, even though each individual bee has really no part in that process.

SPEAKER_01 23:31

That is certainly am I correct in thinking that you're hoping at some point to develop robots that can behave in swarms collectively the way bees do?

SPEAKER_06 23:41

That is certainly what people uh what people and myself included was looking at is is is this even possible? And I think that uh there are there are a lot of in my mind, at least in the current development, these mathematical models that are that uh are currently throughout everyone trying to figure out how these complex behaviors could emerge from simple behaviors, that is really the first step. And then after that, it is I don't think we're going to get to that kind of that kind of capability until the individual agents are actually much more intelligent than they currently are. Yeah.

SPEAKER_00 24:14

So, Colonel Cody, listening to you put all these pieces together, and in particular um with respect to your research about the bees, you know, what I hear you saying is that it's really focusing on looking and observing about how these animals or insects sort of think together in their real-time systems, how they deliberate efficiently to sort of converge or come together to make sort of the optimal solution for a particular circumstance that they have to deal with. I'm just curious, in your research or observations, did you ever come across any circumstances where the optimal solution wasn't achieved or observed or saw sort of any weaknesses or vulnerabilities in the way that in particular bees kind of coalesced around a particular decision point?

SPEAKER_06 25:01

Absolutely. Um I'll start with the the, it's it's funny because I always have to dance between the biological inspiration and then the simulation itself, right? So the work that I did, you know, in his in his book, Tom Sealy proposed a model, a mathematical model. And then that model was initially experimented on with by a gentleman by the name of um Dr. Rana, Giovanni Rana, out of the Sheffield University. Um and so he he proposed uh he has a model and he's been working with this decision-making process as well. And I was working with you know some of his model and then adjusting it to try to improve it. So that model alone, the one, even though it seemed to work well for bees, clearly we didn't have the whole story because um uh if you if you play with it in simulated systems, the bees, because of travel times, would be actually influenced by the environment. So one of the big challenges with uh with these collective systems, and I I dealt with distance between sites, but you can imagine if any one of the options requires more time to assess, then what you have is you have this bias that may either be, you know, it's a bias that would be considered uh a positive bias. And there's several researchers that have that have worked, you know, with Dr. Ana and others, they have this bias in the environment that actually influences uh the decision-making process in a bad way. So the honeybees are relatively immune to not immune to this, but relatively resistant to this problem. Uh, whereas the individual uh agents and certainly the collective of artificial agents, uh, that first model would basically choose the closest site, you know, nine times out of 10, which is not really what we want if we're trying to compare the best one. And so the next thrust uh within this decision-making process was how do we overcome that in a way that still doesn't destroy this behavior? Um, and then you know what kind of delay, you know, possibly you enter distance, which is simplistic. Uh, but one of the things that I explored was uh having the bees adjust their behavior uh given a maximum range, for example. Say the bees know that they're not going to, the artificial agents know they're not going to exceed a certain range from their decision-making location. So a swarm, they would have to, and in knowing that, they would actually augment their behavior to account for that, which is tricky as you can imagine. So the model, mathematical models aren't quite there yet. The honeybees are pretty good at this. The challenge that they have is how do they make how do they pick a site before they run out of internal energy, right? Because they have limited resources in this temporary location that they're living at. But in general, they're able to find and locate uh a new location and move to it pretty reliably. Although there are possibilities for, there are possibilities for failure in the decision-making process. Um, there are other researchers that are looking and finding that just as the simple mathematical model kind of is roughly analogous to a leaky accumulator model for neurons. So it's uh how simple decisions are made within uh even more complex brains, mathematically looks very similar to how input and uh you know recruitment happens within the bees. You know, you have one signal that is increasing in favor among uh among other sensory inputs, uh, and then when it reaches a certain threshold, it fires. And that's similar to reaching a certain threshold for a given site, and then the honeybee is taking off. Um there's uh it's rare, but split decisions are also possible. So you've got a problem where you know assessing the accuracy where the biology may be outperforming um currently the agents that we have uh for uh simulated hives or simulated swarms. Uh and then you have uh the the very low but important risk of um of a group splitting off, you have to guard against as well. Does that get close to what you're talking about, Amit? Great.

SPEAKER_05 28:36

Absolutely.

SPEAKER_01 28:37

Thank you so much. Yeah, doing these studies to see what you can model from swarms into AI, are there certain aspects of bee behavior that you need to put aside, even though it's optimal for the bees? I was thinking of three while you were talking. One is that in a particular temperature, the bees have to remain in a thermal niche and they're burning energy to do that. A second is that they appear to have a division of labor within that swarm, and yet it's not a strict division. In other words, if certain bees are unavailable to perform certain tasks, others can do it. And then there was one you mentioned to me in a previous conversation about certain bees, worker bees, preparing a queen for the flight where they bump into her in order to stop her from eating, in order to get her to lose weight to make her more flight worthy, if I've got that right.

SPEAKER_06 29:34

You absolutely do. It is quite amazing the number of tasks that are happening concurrently with the bees. Now, obviously, although I would have loved to have tried to model all of them, um, the uh it would be you know really uh really difficult to do so. Uh, I'll start with the with the the the one you just mentioned. So all that coordination that I talked about, you notice I didn't mention the queen really once. Uh she is is the queen is not in charge. She's a very important role, a critically important role to play, uh both in promoting good behavior, but also in by releasing pheromones, but also in uh uh, of course, making sure the hive survives. Uh and I say the queen, that's that, you know, there are other queens that get born eventually, which means that the the uh the old queen will eventually have to leave, depending upon uh how the rest of the hive is doing. But yeah, all of these these this task allocation is a uh a problem all on its own. So honeybee hives need water, right? They need they need water collection, they need food collection, they need to move their dead out of the hive, they need to, they need to protect the hive, they need to build the hive, they need to feed their young, all these tasks and regulate temperature, right? Um all these tasks are are critical to make the hive hive survive. It'd be difficult for me to say that I, you know, that you could apply many of those behaviors, and certainly the task allocation behavior makes a lot of sense if you were trying to create a complex artificial swarm where you have some agents doing some, you know, one activity and other agents doing doing another. That may make sense, but I think it only makes sense in a situation where you have um one of the critical features of all these insect colonies is that they have a centralized decision making hub. Um and other researchers to focus have have mentioned have proposed this. And Dr. Rain, as I mentioned before, uh used it in his and I used it in mine, because in order to share information, they have to actually be communicating together. Um, I used a physical space, but as long as there's a way for them to communicate, they have to actually share that information. So the um if the information sharing isn't happening, not only is it hard for them to dance and recruit for foraging sites or uh future nest sites, there's actually also no ability for them to sense which tasks are important. So the task allocation, what I do next matters. So the behavior that you mentioned that I neglected to is this is when the queen is getting ready, when everyone's trying to get ready to go in terms of emitting the swarm, the uh the queen hasn't been flown in quite some time. And so, and so the the scouts, when they start assessing that it's time to move, and that that time to move is task allocated too. It's related to the density of uh bees within the hive. When it gets to a certain population level, it's time to go. And so they start preparing the queen by basically forcing the queen to essentially get flight ready, lose weight. They they bump against her and make her move around and don't let her rest until she's ready to go again. And then when she's ready, that's one of the signals, although there I'm sure there's others, to get them to emit the swarm.

SPEAKER_01 32:35

So you've demonstrated that swarms pursue their own needs and interests and do so, building up from individual behaviors into a collective as a military officer and without taking you into anything that you're actually working on. Theoretically, does a human need to give a drone swarm direction? Does the swarm develop intentions of its own? Or is there some hybrid in the operation as people start working on either military or dual use applications of these swarming behaviors in very small, uh, very large numbered robots?

SPEAKER_06 33:10

Yeah, that is um, of course, that's a great question. Um, in in terms of you know, where does the where does a human and how does a human uh you know exert their control? In terms of, you know, do they need to tell the swarm? Right now, where we're at with current technology in terms of swarming behaviors, it would be very difficult to see. I don't see, I currently don't see any publicly available information about you know swarms that are able to do anything more than move in a coordinated way together. You know, I can keep, I can keep agents, I can give uh uh a direction to um a swarm of UAVs, cutting-edge technology is like, let's get the swarm to go from this location to this location, not run into itself on the way, for example. So that's we're not at the point yet where they're where they're making more complex behaviors. But I think if um in terms of what giving uh giving the swarm direction and giving an intent, I'd say hybrid because I can I can think of you know situations where it would make sense for the swarm to be very tightly controlled, and then others in situations of it's moving away from away from people or away from in a in a uh communication denied or contested environment where uh it needs to operate more on its own, more to more together without direct human control. I think that the you know, so does the human need to give the drone swarm uh direction? Right now, certainly. Right now I don't see anything happening uh where a swarm can do much on it on its uh on its own uh reliably. Um in the future, as they get more complex and are tested more, uh, you know, that could change, but I don't think we're there yet. Sort of uh uh would you like to ask a follow-up?

SPEAKER_01 34:45

I would like No, that that that you hit that one. Thank you.

SPEAKER_06 34:49

Uh a follow-up disclaimer, by the way. I know Charles is about to uh chime in. I apologize, Charles, for cutting you off, but uh I will say uh it it's I'm not working on any of these projects right now. Um and so the uh you know this is uh this is just me keeping taps on on what's been going on out there that you know anyone can look up and applying what I've done for Potter My Research to it. So sorry, Charles, go ahead.

SPEAKER_04 35:10

No, no, no, fine. Good. So I I'm teaching a course here at Dartmouth, and uh we're currently looking at cognitive computing, and of course, autonomous systems or swarms thereof uh has come up. And of course, the students uh wander off in all directions about what you might do with swarms, and of course, there's always a few that start thinking there might be military applications. So if we ever, you know, despite the arguments of whether AI you know, really strong AI is around the corner or around the century, if we beat that, if we're able to do the AI thing and imbue some level of intelligence or at least uh autonomousness in these drones, I mean, let's just imagine. What could you imagine these AI drones could do in a particularly in a military application?

SPEAKER_06 36:00

Sure. Um I always I always like to start with uh with some of the applications that aren't military, um, but I don't mind military ones. So, you know, it's interesting. Some of the colleagues of uh Dr. Raina, you know, they he would he had done a lot of the site selection stuff, uh, a lot of this decision-making process, which is great. And others have worked on things ranging from agricultural. Uh so imagine artificial ants or bees tending large fields in ways that uh that would be difficult for humans to do. So I mean you can imagine the a swarm of you know, harvesters or a swarm of um, and when I say swarms, you know, these don't have to be like tiny little robots, they could be swarming, but larger, large enough to do interesting or important things um with in agriculture from pest control to weeding to you know to harvesting. And that it, I think if you your listeners were to Google and check it out, there's just stuff that is, you know, that people are looking at for that. There's it's been postulated, Marco Dorgo is one of the fathers of the swarm intelligence movement. He's one of the authors of the book uh in the 1990s book, Swarm Intelligence. And he recently talked about, hey, what's the future of swarms? And from a non-military standpoint, he also threw out, in addition to military aspects, you know, search and rescue missions and ecological monitoring, where swarms really are or could be amazing. I can't say they are amazing because they're not there yet. Where they could be amazing, a place where is where you need uh persistent, uh persistent surveillance for any reason, whether that be environmental monitoring, um, and and I'll go in a minute how that could have always obvious implications, at least in the near term, for military engagements. Because you could have a large area covered by a large number of cooperating agents that can monitor, monitor and share information, uh, and even groups of these agents that may perform some of this decision-making process that I mentioned previously that may help them improve, uh, improve the visualization of the battle, of I said battlefield, could be agricultural too, though, right? So uh so yeah, so if from a military aspect, you know, absolutely, your students are your students are right. You know, if you look at sci-fi, uh pop culture is is full of all kinds of references. Anyone who's watched Black Mirror, uh, which is uh has a B episode that uh don't let the kids watch because it's kind of scary. Um, you know, there's lots of horrifying, um, horrifying images out there. Uh what I think about when I think about, and this again, this is just me when I think about it, is uh I think early on we're going to be enabling uh the eyes and ears of land, sea and air forces. So imagine you just you know you have a swarm of agents that are essentially being escorted by a larger agent, whether that be uh on land or sea or air. And there's lots of advantages to just that behavior. It may just be swarming with some very simple behavior, behavioral commands, but the butt that that one agent, whether it be a soldier or uh um or any other um any of the other services, still exerts a fair amount of control because those are you know close proximity, but but the advantages could be significant in terms of pushing out the their own perimeter so that the engagement area is pushed out away from the actual individual um uh military member and uh and also can see a larger area potentially. And you know, of course, if if they are weaponized, uh they'd be very useful not only in traditional traditional attacks, but also if you know engaging other swarm agents, for example. Um again, speculation because we're just not there yet, but that's that's one area. I also roll in the nine separate conversation once, you know, we also talked about the possibility of them being used in uh to surveil urban areas, which is still a hard problem for soldiers. Uh if you have a swarm of you know, bee-like robots, for example, that you let into an urban area just to just to look and then maintain surveillance of an urban area, you know, one of the challenges now is if you clear a building, then you got to continue to watch it forever. All right. But if you leave a swarm of agents there, you can be alerted that something has changed. Um and that's uh that's another important application for it. Um surely you can also imagine, you know, coordinated uh munitions as well, which is uh, you know, obviously more kinetic and and brings up, I'm sure, other questions.

SPEAKER_03 40:09

Warms up one of the things that you're suggested to me is And I got totally unmilitary, but uh one of the challenges we have with recycling is that most of our stuff ends up in the trash dump and it's all mixed up. Much much of it can be recycled if we could only separate it, and but we can't separate it practically. But imagine a large swarm of little robots that aren't very powerful, but just good enough to tell glass from metal, from paper, from food. And then drag pieces to different places. Yeah. And it doesn't really matter that they're slow and stupid, um, because we have plenty of time.

SPEAKER_06 40:57

Yeah. I you know, it's funny you say that, Mark, because because when that my so my advisor is Dr. Julie Adams, uh, she was at Vanderbilt University, but now she's at Oregon State University, and she's continued a lot of this this work, largely focused on how humans interact with those things. But I'm laughing because early on we talked about it, that was precisely the uh the application I was interested in when we first started talking about coordinated movement. I I then moved on to decision making, but I 100% agree. There's other uh research out there that similarly is you know long, long, slow process, but widespread to evaluate infrastructure decay or um or evaluating uh you know eroding pipelines and things like that. Uh, you know, you if you're looking at uh um uh monitoring the infrastructure for whether it be the electrical grid or you know, pipelines or railways or roadways, uh that's a lot of work for people. And uh and you could just you could decentralize it and uh and cover it um without using very expensive, hard to image um uh individual robots.

SPEAKER_01 41:58

One of the I had a couple of questions on challenges to human swarm interaction. And one of them is the way you've been talking, you've taken a piece of swarm behavior and you've studied that. And somebody else has taken a piece of swarm behavior, and and you know, we sometimes talk about you know AI as being like the great elephant that everybody has, you know, a sort of different grasp of. But when you're dealing with something like a swarm, which is acting as a system, are you really able to abstract usefully from a piece of it and ignore the rest of the system?

SPEAKER_06 42:34

One of the larger challenges um with swarm interaction is if I'm if I am piloting a single robot, for example, if I'm a member of a team that's trying to drive a robot, uh I'll have an interface that shows me where that robot is moving and status updates of what that individual robot is doing. That is, of course, important for me to predict what it's going to do next or identify when something's going wrong or know the state of the system is pretty easy because I'm watching it. I think that when you're looking at a collective, or I I'll I'll pause for a moment and explain why I keep using the word collective versus swarm. So to me personally, and this is not this is not held throughout the field, but really you can have swarming behaviors like things like fish schooling or even bees moving together in a group. And to me, that's different than a collective acting as one organism. Uh, and so you know, the honeybee hive is more like a collective organism. So sometimes you'll hear me say collective making a decision. The challenge there is I if I watch one agent and that's all I've got, it's really hard for me to predict what the rest of the swarm is doing, what the rest of the agents are doing. So one of the big challenges with human swarm interaction is how do I improve the how do I shorten the time it takes for a human to look at the swarm and or at least an interface with swarm? What do I need to show the humans so that they can quickly discern what that collective is doing? So in my own research, you know, I was a simulated environment, they were watching, they were watching collectives. Within an environment and uh the humans involved in the study. And uh and you can see that there is a there is a uh significant cognitive effort that occurs when you're trying to figure out how these complex systems and what they're going to be doing. And depending upon the awareness of, you know, if you were to even if you brief the human, I think you can reduce it if the human understands the decision-making process, because then they can more quickly articulate what they think the swarm is doing. But it takes a period of time that's larger than observing a single organism or single agent that's operating on its own, it takes longer to figure out what the swarm is doing. And so another parallel area of research, uh, it's continued by one of my colleagues in this particular problem, but through many other problems in terms of swarming, because in in spatial swarms where they're moving moving together, you still have the same problem. If I can only see part of the swarm, if I see one fish or a group of fish, I don't know what the whole thing is doing. Same with one agent and multiple agents. And so, you know, Dr. Adams and some of the follow-on colleagues, some of the other colleagues, Karina Roundtree, one of them that picked up the work and published, and she was focused primarily on, you know, how do I get to this transparency? What kind of what kind of visualization techniques can I use to make it clear to the operator, to reduce that overhead? Because you can imagine that that trying to predict the one or one robot is hard enough. And then how this robots interacting with other robots and what they might be doing together, even though they're not doing that activity now, uh, that's a lot of overhead. You can get there with interfaces, but uh that imposes their own challenges. Hopefully that that gets closer to some of the challenges that are out there. Uh, there are actually you know others, and we can go into them now or I can wait. But in terms of that all feeds into how much do I trust the system, right? How much confidence do I have in it if I don't know what it's doing at any given point in time? And then if the swarm operates well on its own, how complacent am I if it's done well so far and I have to be able to recognize anomalies?

SPEAKER_01 45:55

I think I think Charles has a couple of follow-up questions he wanted to ask from an earlier topic that I jumped over, and I apologize for that.

SPEAKER_04 46:02

No problem. It's all it's all fascinating. Well, okay, back to my lunatic students. They they really they really took off on this. And so, you know, one one group was saying, well, is it realistic to talk about swarms of you know, military drones or police drones or you know, crowd control, whatever you want to, you know, however you want to arrange it. But is it realistic to talk about those in terms of a human still in the loop to handle C and C or command and control uh up until the time it's obvious what they have to do? Because what you know, the point that I raised in class was the whole goal here is uh fast, light speed. And slowing it down for a sort of mother may I uh it's it's defeating the purpose. And then a follow-up would that to be uh if you're if you're trying to herd cattle, that's one thing. But if you're trying to herd other drones who are acting at the same speed you are, oh golly. Is any of this realistic? Uh I think it is with abstraction, right?

SPEAKER_06 47:14

So I think um you know, it is likely catastrophic um to a swarm if I were to try to take control of any one or even a group of agents within the swarm. It would um because they are making they the interconnections between if you think imagine just a uh three-dimensional cube, a space, you know, uh in the air with 50, you know, UAVs flying around, uh, you know, unmanned aerial vehicles flying around inside uh that space, there aren't any human, there are not even 50 humans, could keep those things from colliding into each other. Um, and it would so you have to rely on the storming algorithm itself. So for individual agents, um, you know, and there's there's many levels of this, and I'll get there. But the first one is for some behaviors like coordinated movement, there are certain aspects of the behavior that I doubt would ever be in control of uh each individual agent's movement. That would be, you know, one of the biggest goals of storm intelligence, and I think you alluded to it, is you know, currently we have you know one-to-one ratio between human and robots, you know, and and that's not even really true. We actually have many humans to one robot in most cases, especially with a sophisticated robot. Um, and clearly, if I have a swarm of 50 agents, I can't possibly afford to have you know hundreds of people watching 50, 50 agents. And so what swarm um behaviors let you do is let you reduce that, uh reduce that ratio. So if I if if my behavior is simple enough and I'm just basically I've got an interface where I'm giving waypoints and I'm saying, okay, this swarm is going to go from point A to point B, it makes sense for me to be in charge because I'm sending swarm uh a swarm in a in a location where I've told it where to go and I've given it a destination. But I actually don't, I may be able to, depending upon how far away I am and what my communication capability is, I may be able to still abort said mission and have them come back. Um, but I'm not telling them, I may not even tell them how to necessarily get there, and certainly not how to how to fly, how the individuals fly and coordinate to get there, um, right? Because that would be very difficult to manage uh those interactions. So if you have, if you imagine a hierarchy, a hierarchy of these behaviors, the mother may I question, I think is more relevant is as you have more complex behaviors that they're capable of. So if I have, if I say if I can if I can assign a mission or a task to a swarm, just have it go take care of it, I'm assigning the task, but not the individual agents and telling them maybe I'm telling them how to get there. Um so that's where the mother may I make sense. If you're talking about, you know, imagine two swarms slamming against each other, for example. So Paul Sharia, of course, in his book Army of Nun talks about one of these experiments. Um, and and everyone should read that book, I in my humble opinion, because I think it does a pretty good discussion about certainly better than I'm going to do about what's out there uh now and what could be happening in the future, but talks about some sort, some basic swarm-on-swarm behaviors, and clearly that kind of you know, a uh a mass of uh agents fighting each other would be would be very difficult for any human to control. Yeah.

SPEAKER_04 50:16

Thank you. That was great.

SPEAKER_01 50:18

Jason, let me go back to the thread that I was on before now on challenges to human-swarm interaction, and then uh a question having to do with why swarms instead of humans. Um you mentioned that it's hard to have a human observing lots of drones at one time. Would it be possible, for example, to have AI tasks with observing and predicting where the drones were going to move so that you wouldn't be relying on the limitations of human cognition to master something that needs to be done that rapidly?

SPEAKER_06 50:53

I think another colleague who's looked at this, uh not a colleague of mine, but a colleague of Dr. Adams, is uh Dr. Mike Goodrich at a BYU. Um one of the, he did a he did some looking at um, you know, how could we how could we better predict what these this coordinate behavior is doing based on number of observations? Like how many observations does it take for me to know what a moving swarm or what formation the swarm is taking, for example. And so you could use, you know, the once you're talking about inputs or sampling of different agents, um, then you are you know you are looking at a possible uh AI problem to help resolve what the state of the organization is. So, for example, uh rather than carefully handcrafting an interface that guesses at uh or tries to articulate what a collective might be doing or a swarm might be doing, um, you could try to train uh train an AI uh to assist in that, um, to first assist in that interface to the human, right? I think what you're getting at is could I delegate to a separate AI agent to control, say, an individual swarm? Uh I see you're nodding, so I assume that's what you're what you're actually getting at. I I'm of course focused on the human-swarm interaction. So um I think that's um there are several several directions that uh that this could go. So if I'm thinking about uh you know a sprawling farm with multiple swarms doing multiple tasks at once, it may make sense for me to have whether it be self-organized or or separate agents that are keeping tabs on what the overall swarm is doing. We are uh I am, I won't say I'm in the sci-fi realm, but I'm definitely not in the near term in terms of capabilities. But I think it's not out, I think it's then the realm of possible. The challenge, there's lots of challenges that come up though. Anytime you see that kind of observation power is that challenges with AI everywhere else, which is uh, you know, what's the training data look like? How well do we test it operating in real environments? You know, how much do I do I trust without seeing what that AI is feeding back about the swarm? And honestly, I don't have a good answer for that. And I certainly wouldn't trust the system without uh without seeing uh without seeing extensively tested in the actual environment it would be operating in uh at this point, because they just aren't the behaviors aren't resilient enough for me to throw it in an unknown environment and expect it to act well.

SPEAKER_01 53:11

Yeah, so and I know this is gonna sound rather strange at the end near the end of a discussion of swarms and of insects and bees written and their behavior, but um why do you think it's more useful to look at swarming insects as a model for AI rather than collective human behaviors? I mean, humans don't necessarily swarm, but there are crowds, there's crowd behavior, there's mobs, there's mob behavior. Um maybe they're not as intelligent or as functional, and maybe that's why you want to go to swarms. But could you tell us why you found that it was more useful to go to insects for the model rather than humans?

SPEAKER_06 53:54

So one of the uh it's interesting because uh we were one of the challenges we use in human human swarms, right? Yeah, yeah, I I like to I had a bit of an existential crisis, honestly, when I started studying this, because when you start thinking about superorganisms, uh then you start thinking about cities as colonies and you start thinking about you know our own movement behaviors and how you know a lot of what we do is also self-organized, uh, and it makes sense. Humans inner have an incredibly complex way of interacting. Um, and so I think that um, you know, certainly there's been studies not in this, not in this area, although there's there there are some where there, you know, movement movement within uh individuals, sorry, movement like a mob moving, for example, or or if you look at anyone who's you know watched the New York Subway or moved up and down the street, you know, we have these, we have these little behaviors that we kind of take for granted that that dictate how we move in with relation to each other. You know, different cultures have different you know social space that they're comfortable with. And and that's you know, when you think about it, it's kind of like swarming. I in I I joke when I walk with my colleagues, if someone bumps into us and I'm like, yep, swarming, I'll give them broke, you know, because it's that's what it is. But part of the reason why insects are valuable is because typically their interactions are more, are a bit more predictable, they're a little bit more regimented, meaning that you know, a honey bee swim, this honeybee call uh hive and this honeybee hive and this one, they all typically have similar behaviors uh that they've evolved over time that are that are observable and and uh and repeatable in experiments. And that that may not be true of humans because there's so many ways of communicating and different ways of interacting. Um, one of our big strengths is the fact that we can actually change how we interact with each other. Um, whereas whereas a single honeybee colony, uh, although there are individual changes between the members, uh individual differences between the members, largely interact um similar manners. So I would love to pull that thread of different individuals, but uh we can get there if we have time. You know, humans.

SPEAKER_01 55:59

Well, and I I was thinking of some of the ones we've talked about in the past, the behavior of people on 9-11 where everybody started walking north, or the chaotic behavior in collective, if you will, of refugees from a war zone in a train station trying to escape.

SPEAKER_06 56:16

Certainly interact, um, and humans certainly swarm. I I personally believe, and this is not backed up by any research, and so belief is a bit of a strong word, but when you think about it, you know, swarming can kind of happen at multiple levels. So the honeybees, right, they have um they have these swarming behaviors, but they also have this swarming of intelligent of uh of information sharing, right? So I I have a uh one of the members of of my PhD committee actually uh I found out as as we were going through it, had also studied the honeybee decision-making process, but in terms of social belief propagation, which is fascinating, you know, obviously looking at insects, but also looking at how ideas propagate. Yes. Um, and the the insect gossiping models that are out there, how they share information. You could say, I wouldn't say it's exact. I'm not I'm not calling us bugs, but I would say that it's interesting. That's how you know insect colonies share information by essentially gossiping. Um and the interaction rate and the the which agents interact, you know, clearly affects how information makes its way through the colony. Um, as far as you know, I'm not aware, I'm not aware of any swarm intelligence studies that use humans, partly because humans are so complex. And I think that really is what it comes down to. I think that's a factor. I think you could do pedestrian movement studies, you could do optical flow analysis of you know, camera feeds of people moving and come up with similar, similar movement and behaviors. You know, just walking down the street, um, you know, in uh in America and uh in New York City, for example, there's kind of some unspoken rules about how we behave. And uh and I'm pretty sure you could model those too. Uh the difference is I think the differences between us are more pronounced because we are more complex and have have longer lives and more learned behaviors than say uh you know, three to five-week old bee.

SPEAKER_04 58:04

So I think So I guess it's time to wrap up, but let's think, let's look in the future, future directions of inquiry. Where do you think the if we were looking for the beef, so to speak, where would we look?

SPEAKER_06 58:16

That's uh I think I've indicated before, you know, first of all, the the the handcrafted mathematical models, which are valuable for trying to figure out how these behaviors might work, are not useful in dynamic environments. We definitely have to definitely have to figure out ways of learning those behaviors, uh, whether that be some kind of genetic algorithms or you know, um, you know, doing deep learning, trying to figure out, trying to solve the problem of how do I learn an interaction behavior in uh in a dynamic that worked for at least several different types of similar environments uh in a way that will result in the desired collective behavior is that is a significantly different difficult problem. Marco Dorgo proposed this in his recent uh his recent paper where he projects forward a little bit. And I totally agree with him. In the uh the we there was a swarms swarms conference in 2017 in Kyoto. Um and there, you know, one of the things that we all talked about was we talked about, and Marco Dorgo's there, as was Nigel Franks, who I I have to make sure I mentioned because he's a uh he's he's done a lot of work in swarm intelligence. Heterogeneity uh matters, even in insect colonies. And part of what makes these colonies robust and dynamic are is the differences. Even though it'd be easy to say they're all the same, it's it's the differences. So now you've got the problem of okay, can I make a swarm of agents, a swarm of robots, where they have these individual differences that, as a collective, improves their ability to make better decisions or have better behaviors? A hilarious. I'll leave you with with one that was great. I was listening to a study about lazy ants. And uh yeah, some ants are lazy, they're totally lazy, and they they don't they don't work when everybody else is working. And what is the point of these lazy ants? Just eating food. Come to find out they're a latent, uh, they're a latent quick reaction force. Uh, they only get engaged when everybody else is tired. And uh, when you think about it, that's a that's a fantastically evolved uh tactic for keeping the colony alive in times of distress. Uh thank you.

SPEAKER_04 1:00:12

That's a great, great finish. It also reminds me of teenagers. But uh, but with that, uh wow, this is fun. Uh it's really cool to get to hang out with you again and to to hear what you've been doing since the last time I ran ran into you. Uh, thank you so much for your time and for doing this work and uh answering our crazy questions. This has been wonderful. Thanks you very much.

SPEAKER_06 1:00:37

Thank you very much. I enjoyed it.

SPEAKER_01 1:00:45

We thank the business law section of the American Bar Association for their generous sponsorship of the production of this podcast. We welcome questions and comments from listeners. Send email to comments at mindthegapdialogues.com. We read all comments and questions and will try to respond in the letters section of a future episode. If you are writing about a particular episode, please do mention a specific episode number. Please also do include pronunciation tips to help us properly say your name when we reply in a subsequent episode. See you next time on Mind the Gap Dialogues on AI.

SPEAKER_02 1:01:46

Thank you for listening to the AVA Business Law Sections Podcast Series to the extent that the section offers a robust collection of content. To explore more about this topic or to learn about joining the section, visit ambar.org slash bizlaw. That's B I Z L A Well.