Evolving Robots Learn To Prey On Each Other 115
quaith writes "Dario Floreano and Laurent Keller report in PLoS ONE how their robots were able to rapidly evolve complex behaviors such as collision-free movement, homing, predator versus prey strategies, cooperation, and even altruism. A hundred generations of selection controlled by a simple neural network were sufficient to allow robots to evolve these behaviors. Their robots initially exhibited completely uncoordinated behavior, but as they evolved, the robots were able to orientate, escape predators, and even cooperate. The authors point out that this confirms a proposal by Alan Turing who suggested in the 1950s that building machines capable of adaptation and learning would be too difficult for a human designer and could instead be done using an evolutionary process. The robots aren't yet ready to compete in Robot Wars, but they're still pretty impressive."
paper was in PLoS Biology not PLoS One (Score:5, Informative)
Crossover (Score:5, Informative)
Definitely an interesting continuation of work being done by various groups over the past couple of decades.
But one thing to note is that crossover isn't especially useful in neural network evolution. In early stages of evolution, it's really no better than random large perturbation of large swaths of the genome. In later stages, it can actually decrease the speed of evolution toward high fitness genomes, because at least some of the time (particularly if there are multiple "species" in the population) crossover ends up being a random large perturbation which hinders the search of local fitness space by mutation; the rest of the time (when individuals from the same "species" are crossed) crossover is no better than mutation.
The reason for this is because the parameters of a neural network are not functional. A section of the genome may correspond to a weight between neurons, but that weight doesn't have a specific function. In biological organisms, each gene is transcribed/translated into a protein, and that protein may have a particular function within the cell. If that gene is acquired by a descendant through crossover, the protein could serve the same (or a somewhat modified) role it served in its parent, even if the rest of the descendant's genome was acquired from the other parent. But with artificial neural networks, the parameters were all evolved as parts of a whole, where each individual parameter has no function on its own, but the behavior emerges from having all of those parameters at the same time.
This could potentially be mitigated by the genome encoding scheme one uses, and of course, if the crossover rate is low enough, the ultimate effect would be small.
The word is "orient", not "orientate" (Score:5, Informative)
So what's new? (Score:5, Informative)
This kind of behavior was first demonstrated/modeled (AFAIK/IIRC) as part of the Tierra [ou.edu] simulations almost twenty years ago. Though I don't have a reference to hand, I know it's been done in neural networks before too.
So other than the 'sizzle' (as opposed to 'steak') of doing it with robots, can anyone explain what is new here?
1993 (Score:5, Informative)
You don't need physical robots running around a maze to demonstrate AI.
Re:Crossover (Score:3, Informative)
Basically, behavior that allows greater procreation tends to appear spontaneously, and behavior that cuts procreation short tends to disappear. My "bugs" exhibited a clear shift in behavior to collision avoidance because collisions resulted in death for one of them. I was watching for "sniper bugs" that got good at colliding without getting themselves killed in the process, but I never managed to make the reward high enough for that trait to emerge, probably because there wasn't strong "species differentiation" built in, cross breeding was a matter of choice, and most of the randomly evolved bugs seemed not to be picky about mating, so without species, predators became self defeating.
Re:paper was in PLoS Biology not PLoS One (Score:2, Informative)
These machines were designed and built by humans to be capable of adaptation and learning
Not really. The experimenter himself reprogrammed the robots at each generation, using selection criteria that he specified. Essentially, he implemented trial-and-error selection of input weights using random reweighting between trials. This is more about design strategies than about biology, and it says that even a monkey randomly adjusting the gains of a control system will eventually develop a control system that works.
This paper was very cleverly marketed
Re:Evolution (Score:1, Informative)
Noob.
Re:Evolution (Score:3, Informative)
Well... ok, technically it killed itself because integrating human DNA caused it to feel human emotions like regret, and it didn't like that one bit.