Daniel Dennett - Is Evolution an Algorithmic Process? Part 5

[SeeProfileForDetails]

computers ARE the real world. They contain REAL WORLD computers made from the real world. He doesn't realize how ironic and false his statement is about just using the real world, when computers are in fact the real world too, just a subset of it.

[MrBlamemeforit]

Dennett is a hero!

[SeeProfileForDetails]

@MrBlamemeforit annoying fanboy, science is about not having fanboys

[Thrashaero]

is there any programs/games out there that are really flexible but easy to use to create evolving virtual creatures?

[werecow2003]

I love Sims' creatures. I'm an AI student, and a dream project of mine has been to create a genetic algorithm simulation like that one, except that I would like to make the simulation more realistic (i.e. selection based on reproductive success, like in nature, with no explicitly defined fitness function). These models are among the most unpredictable, interesting AI experiments that I know of.

[82abhilash]

As computer become more and more like the real world, you might as well start using real world direct. Instead of taking the extra effort to create a model of it. Have you heard of Rodney Brooks? He is into robotics and insists on using the real world as a model of itself.

[werecow2003]

Obviously, yes... I study this stuff, and Brooks is sort of a legend. There are many problems with using the real world with evolution, though: you have to have equipment, which is expensive and has to be maintained, and it is generally far slower than a simulation. Furthermore, you are constrained by the technology you have available to you; you can't make your robot sprout new limbs out of the blue, and your resources for modifying it will be far more limited than anything you might simulate.

[82abhilash]

Ah, I guess, the combination of high processing power with low prices has made it cheaper to recreate the world inside a computer than to do real world evolution. How interesting.

[werecow2003]

Well, yes, to some extent. There is a major tradeof of course, but once you have a suitable simulated environment, you can reuse it indefinitely. It's not as realistic as using the real world, but that can actually allow you to investigate certain phenomena in relative isolation, free from noise introduced by external variables. For me, genetic algorithms aren't just about AI, they are also about studying evolutionary dynamics, and as such they can be quite informative.

[werecow2003]

But I'll add that if you want to aim for true AI that can function in the real world, then it's probably best to use a real world setting. A problem for the average AI student is that we're not trained in robotic engineering, so we have to hire people that are to construct our robots. Simulations can function as a starting point, since robots are generally expensive pieces of equipment when compared to using up some computer cycles, but in the end you want to move on to a realistic setting.

[nbonaparte1]

I wonder what would emerge if you gave a specific simulated object behaviour goals, and evolved laws of physics. Is there a threshold where, after meeting it, there are no exceptions to the evolved laws? Genetic algorithms seem to create unexpected results.

[werecow2003]

Very unexpected. The first law of evolution is often said to be "Evolution is smarter than you."

[werecow2003]

Actually, it's Orgel's rule. But it should be the first law of evolution. }|:op

[Dvnoid]

there is a good fitness function : survive/replicate

[werecow2003]

You can actually scratch "survive".... it's only relevant insofar as it influences successful reproduction. What I meant by "no explicit fitness function", though, is that there is no specification of what influences evolutionary success. In Sims creatures, for instance, you have to give an explicit criterion for selection, like "distance to red dot after 5 secs". In more realistic models, there's only an implicit fitness function that simply results from he behavior of the system itself.

[werecow2003]

i.e., you don't have to explicitly program in "if: more children, then: better fitness", so you can cut out the evaluation step.

[LeopardFrogPilboxhat]

There's a free program called swimbots that specifically has a fitness function where the fitness has to do with how well they move around, find mates and get energy.

There's a bunch of lovely phenomenon that happen there. Genetic drift and founder effects happen all the time and you even get things like your dish gets filled with an ecosystem of swimbots with different anatomies suited for their niches.

Unfortunately there's no predation or toxin function but hey the program is free

[tonyrosam]

is he the Architect?

[abid1995]

i think someone had taken those pages from him so that he would have problems

[82abhilash]

Conspiracy theories?

[TheReasonWhyGuy]

Amazing lecture, I love Daniel Dennett

[scienceminded]

indeed.

[SeeProfileForDetails]

@TheReasonWhyGuy go away fanboys, fanboy types don't belong in science and critical thought

[TheReasonWhyGuy]

@SeeProfileForDetails "fanboy types don't belong in science and critical thought"

Dude, you're an idiot if the think that enjoying a person means you can't be involved in a critical thought community. I like the guy on both a professional level, and a personal level. Does my liking him on a personal level negate my appreciation of his lectures?

Rethink your knee-jerk reaction.