Robots evolve to deceive one another

Blogging on Peer-Reviewed ResearchIn a Swiss laboratory, a group of ten robots is competing for food. Prowling around a small arena, the machines are part of an innovative study looking at the evolution of communication, from engineers Sara Mitri and Dario Floreano and evolutionary biologist Laurent Keller.

They programmed robots with the task of finding a "food source" indicated by a light-coloured ring at one end of the arena, which they could "see" at close range with downward-facing sensors. The other end of the arena, labelled with a darker ring was "poisoned". The bots get points based on how much time they spend near food or poison, which indicates how successful they are at their artificial lives.

They can also talk to one another. Each can produce a blue light that others can detect with cameras and that can give away the position of the food because of the flashing robots congregating nearby. In short, the blue light carries information, and after a few generations, the robots quickly evolved the ability to conceal that information and deceive one another.


Their evolution was made possible because each one was powered by an artificial neural network controlled by a binary "genome". The network consisted of 11 neurons that were connected to the robot's sensors and 3 that controlled its two tracks and its blue light. The neurons were linked via 33 connections - synpases - and the strength of these connections was each controlled by a single 8-bit gene. In total, each robot's 264-bit genome determines how it reacts to information gleaned from its senses.

In the experiment, each round consisted of 100 groups of 10 robots, each competing for food in a separate arena. The 200 robots with the highest scores - the fittest of the population - "survived" to the next round. Their 33 genes were randomly mutated (with a 1 in 100 chance that any bit with change) and the robots were "mated" with each other to shuffle their genomes. The result was a new generation of robots, whose behaviour was inherited from the most successful representatives of the previous cohort.

In their initial experiments, the robots produced blue light at random. Even so, as the robots became better at finding food, the light became more and more informative and the bots became increasingly drawn to it after just 9 generations.

But as it is for real animals, it's not always in the robots' best interests to communicate the location of food. The red ring only has space for 8 robots, so not if every bot turned up, they had to physically shove each other for feeding rights. The effects of this competition became clear when Mitri, Floreano and Keller allowed the emission of blue light to evolve along with the rest of the robots' behaviour.

As before, they shone randomly at first and as they started to crowd round the food, their lights increasingly gave away its presence. And with that, the robots became more secretive. By the 50th generation, they became much less likely to shine near the food than elsewhere in the arena, and the light became a much poorer source of information that was much less attractive to the robots

Nonetheless, the light never became completely useless. As the bots became more savvy in their illuminations and relied less and less on the lights, individuals that actually did shine near food pay a far shallower price for it. Because of that, the evolutionary pressure to keep others in the dark was lower and the information provided by the lights was never truly suppressed.


This also meant that the robots were incredibly varied in their behaviour. With the yoke of natural selection relaxed, processes like genetic drift - where genes pick up changes randomly - were free to produce more genetic diversity and more varied behaviour. After around 500 generations of evolution, around 60% of the robots never emitted light near food, but around 10% of them did so most of the time. Some robots were slightly attracted to the blue light, but a third were strongly drawn to it and another third were actually repulsed.

Mitri, Floreano and Keller think that similar processes are at work in nature. When animals move, forage or generally go about their lives, they provide inadvertent cues that can signal information to other individuals. If that creates a conflict of interest, natural selection will favour individuals that can suppress or tweak that information, be it through stealth, camouflage, jamming or flat-out lies. As in the robot experiment, these processes could help to explain the huge variety of deceptive strategies in the natural world.

Reference: PNAS 10.1073/pnas.0903152106

More on robots:

i-77217d2c5311c2be408065c3c076b83e-Twitter.jpg i-3a7f588680ea1320f197adb2d285d99f-RSS.jpg

More like this

Oh, but they're all still robots so this experiment proves nothing! :P

I think I saw a similar paper a few years ago - may have been from the same group. It wasn't quite so complex - I'm not sure if they reported any findings on genetic drift in the earlier one.

In any case - I LOVE THIS PAPER - thanks for writing about it.

...and I, for one, would like to welcome our new mechanical overlords.

(C'mon, you knew somebody was going to say it.)

So it occurs to me wonder if, under the right conditions, the robots could evolve in just the opposite direction, i.e. broadcast the position of food to each other as brightly as possible in order to alert other members of the species (other carriers of the same 'selfish gene') to the food source.

Certainly that would not happen in anything like the current experiment, where survival of the gene pool itself is not in doubt. But perhaps if there were competing populations of robots, some with blue light, some with red light, some with yellow, etc., and each "species" can only "see" the same color of light that it emits. The initial number of each species would be the same, but would vary in generation since you still are only keeping the top 200 (or whatever) scoring robots, regardless of species.

Assuming one of the species was able to develop a sufficient population of "always broadcast when you find food" robots, then conceivably that species could begin to outperform the other species. Though they would tend to crowd each other out,

Perhaps this is all fanciful, since the selective pressures on the individual would be too high to ever be outweighed by selective pressures on the species. I can almost see it happening, though, with some type of Mendelian inheritance process, with the "gene" for "always broadcast when you find food" being recessive.Then the gene can have a clear benefit for individuals, as long as they are carriers -- carriers of the gene are still free to deceive other robots when they find a food source, but they are more likely to be surrounded by robots who will alert them to a food source when found.

All this is way beyond the scope of the (totally awesome) experiment. I'm just sorta thinking out loud...

I suspect my toaster is lying to me, and now I'm worried that my toast is poison.

But I still like James Sweet's comment that it would be interesting to see if you could create conditions where their would be cooperation between groups of robots.
What I find kind of funny is that they choose to use actual physical robots instead of a total simulation. What is the advantage of using robots other than the EXXTREME cool factor?

But I still like James Sweet's comment that it would be interesting to see if you could create conditions where their would be cooperation between groups of robots.

I ended up writing a whole blog post about it.Although...

Ed, could you clarify the following statement from your summary?

After around 500 generations of evolution, around 60% of the robots never emitted light near food, but around 10% of them did so most of the time.

The 10% that emitted light near food "most of the time" -- was this more frequent than random? If so, then I guess they already had evolved a small cooperative sub-population...?

I would find that pretty hard to believe because, as I described both in my comment here and in more detail on my blog, I have a hard time seeing an evolutionary benefit to the "cooperatives" under the present experimental setup. Because the population size is fixed from generation to generation, there doesn't seem to be any individual benefit to altruism. If the population size were variable, e.g. by dividing the overall population into species as in my suggestion, then at least there is some benefit to the altruism (increases proliferation of the species as a whole, thereby giving cooperative individuals more chances to mate), although it's hard to say whether that benefit would outweigh the obvious detrimental effect of having to compete for overcrowding every single time you find food.

But hey, I'm a computer engineer, not an evolutionary biologist, so what do I know...

What I find kind of funny is that they choose to use actual physical robots instead of a total simulation. What is the advantage of using robots other than the EXXTREME cool factor?

heh, I mentioned the same thing on my blog...

I assume that the researchers sort of randomly mated successful genomes together--has anybody figured out a way to add sexual selection to this? That way mating preferences could also evolve, and some robots might decide they only liked the shiny ones. Or would that matter?

ambivalent academic: I also remember reading about the same thing two years ago. Probably because I'm not a science guy and not attuned to significant details, but it seems like they had a pretty good handle on all this stuff two and a half years ago. At the risk of going off topic, what takes so long for a science paper to go from the seemingly complete status two years ago to the complete version today?

Here's a link to a story about the research they were doing two years ago:

Hmmm - I accidentally deleted a lot of comments, which have now been restored (hence the suspiciously tight temporal clustering). To answer points I previously dealt with:

- Rob, checking the paper again, it looks like the experiment was a mix of real robots and simulations. Presumably the simulations were used to scale the experiment up to the 100 groups of 10 robots.

- James, 11.2% of the robots emitted light more than 50% of the time - so the population includes those who are doing it by chance, but also some that were always emitting light near food.

- Matt, I hadn't actually seen the previous reports before writing this piece. They're two separate papers though published by the same authors, one in Current Biology two years ago and one in PNAS now. Will check to see what's progressed since then

- Everyone else - the experiments are still in their infancy, so no plans yet on looking at group selection, sexual selection or so forth. But please, keep your wishlist coming.

Wouldn't it be more realistic if, instead of choosing the 200 best robots, you choose all the robots whos number of points exceed a predefined limit? This limit would represent death. Of course then you should let the number of point slowly dwindle, forcing the robots to eat good food.
This, perhaps, more closely represents the actual situation of all species, wanting to eat to prevent them from dying (wanting to score, to prevent the number of points falling beneath the limit) and only the living mating and producing offspring.

The man most directly responsible is Miles Bennet Dyson.

By bobmighty (not verified) on 21 Aug 2009 #permalink

we must kill miles dyson.

I can't seem to find thae actual scientific paper:
* The DOI given (10.1073/pnas.0903152106) doesn't resolve
( says it doesn't exist).
* google scholar "Mitri Floreano Keller robot evolve" finds
other papers by the same authors, but nothing in PNAS in 2009.
* The (very slow!) PNAS website's search engine finds no articles
with authors Mitra and Keller, Floreano and Keller, or Mitra
and Floreano (I searched on pairs of authors, rather than all 3,
in case one of the names was misspelled).

Can anyone provide a citation, correct DOI, or (best of all)
online link to the paper?

By Jonathan Thornburg (not verified) on 22 Aug 2009 #permalink

It's the right DOI. Try again in a week.

PNAS, really annoyingly, often leave a gap of up to two weeks between lifting the press embargo on their papers and actually bothering to publish the damn things.