Swarm Robots Evolve Deception
Writing by Evan Ackerman on Wednesday, 19 of August , 2009 at 3:30 am
In a mere 50 virtual generations, swarm bots (remember them?) using genetic software evolved the capacity to lie to other robots about the location of a source of food. Initially, the robots were programmed as a group to search for an object that represented food, and they gradually learned to emit a blue light when they found the food to show other robots where it was. Researchers at EPFL in Switzerland evolved and mixed the programming of the most successful foragers until they had a bunch of robots who were very good at finding food, and then gave the virtual genes of each individual robot control over their blue light that signified food.
You might expect that the robots would learn not to signal when they found the food to reduce competition, which is passive deception, but they also evolved an actively deceptive behavior, where some robots would deliberately travel away from the food and signal their blue light, drawing other robots in the wrong direction. Crafty little buggers. Interestingly, this deceptive behavior didn’t make much of a difference to the overall fitness of the group strategy of following blue lights… Some robots always tell the truth with their blue lights, which means it’s always advantageous for a clueless robot to follow a blue light as opposed to just wandering randomly.
So why do some robots keep telling the truth if deception can effectively lure other robots away from the food? It’s fairly simple, as I understand it… If all of the robots are deceivers, any new robot will quickly learn that avoiding blue lights is the best way to find food. And in that case, any robot that starts signaling its blue light when it does find food (through a “genetic mutation” in its software) will increase its own fitness by repelling other robots from the food it finds. As it passes this behavior on to its virtual children, there will be more and more truthful robots until it once again becomes more advantageous to be deceptive.
There are, however, populations of truthful and deceptive robots such that the combination of behaviors reaches a stable point. In this particular experiment, the stable evolutionary endpoint (after 500 generations) was that 60% of the robots were deceivers and 10% told the truth. Furthermore, about a third of the robots were slightly attracted to blue lights, another third were strongly attracted, and the final third avoided them completely. This type of experiment, its progression, and the results are particularly fascinating to me because the robots are behaving and evolving in much the same way as populations of animals do. Examples of both altruism and tactical deception can be found in many different species of animals as well as (of course) in humans, but these little robots offer a unique opportunity to study (and tweak) the evolution of behavior in real time.
[ EPFL ] VIA [ Not Rocket Science ]
Comments (10)
Category: Artificial Intelligence, Biorobotics, Research
- Add this post to
- Del.icio.us -
- Digg
Comment by geraldt
Made Wednesday, 19 of August , 2009 at 7:25 am
I wonder why for experiments like this, they bother with the swarm of little robots? Couldn’t it all be simulated in software inside one cheap desktop computer? Oh, cute little things that run around the floor with coloured lights – I guess I can see the attraction.
Comment by Evan Ackerman
Made Wednesday, 19 of August , 2009 at 10:25 am
From the article: “Furthermore, we show that the use of realistic models can lead to dynamics (such as the stability of deceptive communication) that one would not observe in simplified mathematical or game-theoretical models.”
You could argue that a complex enough software model would display the same behaviors, but little things like whether or not a bot is quite close enough to another bot to see a signal light varies slightly from robot to robot, which introduces a little bit of variation and randomness (realistic factors) to the study, and different and/or more complex behaviors can emerge from things like that.
Comment by ME
Made Monday, 21 of September , 2009 at 5:47 pm
I think the part your both forgetting, and the article left out…
These robots are connected to rat brains. There are light rings on the floor, and if a robot is near one, the bot triggers the pleasure and food sensors in the brain making that a good spot to be. And of course various senors are hooked on the brain to do this. This is a merging of robotics and animal. AKA bioengineering.
This of course you can’t make a software program do. One could argue that they should have just simulated this by have the brains control a virtual character… but the purpose is also proof of concept…
Comment by Daniel
Made Tuesday, 22 of September , 2009 at 9:06 am
So then this article is a deception, because Robots have nothing literally to do with it, if their parent was a rat.
Altruism is a characteristic that does not favor the individual. so how is it that altruistic characteristics are found in man when there is no advantage to their long term survival?
Comment by Blake
Made Wednesday, 23 of September , 2009 at 10:58 pm
@Daniel:
Altruism does favor the individual though. It can create an system of reciprocity. If the robots(or anything really with a ’social’ function) was able to inform one or more of the group it existed within of the presence of food/resources then this behavior would be reciprocated to it at some point. By working as a group searching for and sharing food the group increases its efficiency. This idea goes even further to form the basis of human rights. Wikipedia has a decent article on the ethic of reciprocity if you want to look.
The system presented with the robots is perhaps too simple to show this.
Comment by BobOmb
Made Saturday, 24 of October , 2009 at 3:22 am
@ME and Daniel
Rat brains? where did you get that from? There’s no mention of rat brains any where in this article or any of the previous articles. It mentions Linux in the earlier swarm bot articles, nothing about rat brains. Checkout http://www.swarm-bots.org/ if you don’t believe me.
Comment by evotech
Made Monday, 16 of November , 2009 at 6:29 am
And so, the Autobots and the Decepticons where created
Comment by matt m
Made Tuesday, 8 of December , 2009 at 9:10 pm
the evolution of the bot behavior was most likely carried out inside a simulator.
there exist several retail and open source simulation environments for robot populations that can even model things like camera/sensor noise and physics (i.e. collision detection). there are even ways to simulate interference on communications links (radio, optical, etc).
of course, nothing beats using some of your grant money to buy some swarm bots and watch the cute little guys eke out their meager existence, not to mention it looks good in front of the grant committee.
Comment by Gordon
Made Wednesday, 9 of December , 2009 at 6:30 pm
@BobOmb
Linux is powered by rat brains.
Comment by Dennis Sisterson
Made Thursday, 10 of December , 2009 at 4:44 am
Daniel: (if you haven’t found this out by now!) Man has a gene for altruism not becuase if favours the indivudual human, but because it favours the altruism gene. In other words, if the gene exists in a group of humans, it will tend to survive even if the individual who acts on it – thus aiding the survival of the other hosts – dies as a result.