0534: "Genetic Algorithms"

[Wolfkeeper]

The really funny thing is, genetic algorithms really are evil little critters. They really do have no boundaries or ethics and will happily abuse whatever you give them to play with, to a degree that is not funny. But on the upside, they're super-dumb and really slow, so I've never had to do more than beat my printer to death, to date.

[potatolicious]

Randall you fool! If you make the number too large it will overflow and wrap when you start summing up your navigation cost! Then if you minimize the function...

[Beacons!]

At last! A comic I don't understand, haven't seen that in a while. Xkcd is back on form.

[martinw89]

Randall you fool! If you make the number too large it will overflow and wrap when you start summing up your navigation cost! Then if you minimize the function...

Python's ints/floats can be of arbitrary size

[irfan9727]

http://www.wreck.devisland.net/ga/ interesting generic algorithm written in flash.
(from the author: "This is a GA I wrote to design a little car for a specific terrain. It runs in real-time in Flash. The fitness function is the distance travelled before the red circles hit the ground, or time runs out. The degrees of freedom are the size and inital positions of the four circles, and length, spring constant and damping of the eight springs. The graph shows the "mean" and "best" fitness.")

I almost died in laughter after reading this comic.

[Thrice Great]

I have clue what genetic algorithms are, but I have seen all the Terminator movies.

[Teutates]

I really expected more "get out of my head, randall"-answers

Happened to me, I was just starting a little GA-project myself when I saw the comic
The only time I followed through this sort of project was in AI class, solving the TSP problem (works pretty well to get a good solution fast, though you shouldn't expect to get a perfect one). Recently this guy made GAs popular again.

Genetic programming is also very interesting (a subset of GA, your specimens are programs), but I never figured out what kind of language and virtual machine would be best... would be the perfect breeding ground for skynet though

[Bunsen]

Will you believe I got a mid-air collision with you because I posted the same god damn thing?! Star Control 2 reprazzenT!

I think that just means we're predictable.
PRIORITY OVER-RIDE. NEW BEHAVIOR DICTATED.
MUST BREAK TARGET INTO COMPONENT COMPOUNDS.

[Ptolom]

Will you believe I got a mid-air collision with you because I posted the same god damn thing?! Star Control 2 reprazzenT!

I think that just means we're predictable.
PRIORITY OVER-RIDE. NEW BEHAVIOR DICTATED.
MUST BREAK TARGET INTO COMPONENT COMPOUNDS.

Heh, your problems have only just begun. (Cue sinister Ur-Quan theme music)

[frogger626]

Uh, guys? I checked the comic about 6:30 central this morning, and it was up fine. Now, I try to find it at 7:15 (again, central) and it's gone. Thread's here, but the comic is gone.

What's going on?

[Chalnoth]

Love the comic!

Unfortunately, the reality is, how could you write an algorithm to determine whether or not the code had become skynet?

[GreaterSteven]

Normally I disagree with xkcd overrated (the post isn't even up yet, but I know what it's going to say).

That's like the 80th terminator reference in the comic.

I don't mind using code all the time, it offers fantastic humor opportunities. But more terminator references?

[phillipsjk]

I find the Terminator series of movies believable, save for the time-travel aspects. I don't think we will be able to "write" a "true AI", It will be an accident. With the current state of computing, I envision a cross-platform self-modifying worm that runs using machine code. It would be able to bypass OS checks by way of deliberate or accidental back-doors on the ubiquitous x86 chips. That is to suggest: hardware monoculture coupled with hardware bugs may doom us all!

That said, I was never worried about genetic algorithms. It is my understanding that they always have a well-defined problem and run in a sandbox. Any AI facilitating code will not be optimal for the problem at hand (For a large subset of problems anyway).

...The colleague was explaining how he made a multi-layered GA so that the genetic algorithm itself is optimised by a genetic algorithm. This leads to some interesting results. It also caused me to mentally paraphrase, "Yo dawg, I heard you like genetic algorithms, so I put a genetic algorithm in your genetic algorithm so you can evolve while you evolve."...

Now, I did not think of that. The description is not clear, so it may or may not throw my assumptions out the window. I presume that the outer layer has the properties I described.

[canis.uk]

Closer to the real topic, all efforts as described in this comic are wasted if you type 'import skynet'.
Although, if Skynet is sitting in a module on your computer ready to be imported, you're already doing it wrong...

Just in case, I'd recommend:

Code: Select all

from __future__ import kyle_reese


[abstractpotato]

HAHAHHAHAHAHAHHAHAHA!
Brilliant!

[vince9663]

As an extra safeguard, you might also want to add:

"ProtectJohnConnorCost = 1"

[6453893]

For all those who complained xkcd had lost its nerdiness; this one was completely inaccessible to me.

[Sprocket]

Oh no Randall, are you going to make me start watching Sarah Connor Chronicles again...?

[dracolytch]

Dear Randall,
From someone who's actually programmed a few of these, I say unto you: Spoilsport.

~D

[podbaydoor]

*watches as the comic flutters by over her head, beautiful little thing*

[dracolytch]

Love the comic!

Unfortunately, the reality is, how could you write an algorithm to determine whether or not the code had become skynet?

Well, the pseudocode would look something like this...

Code: Select all

foreach (entity e in population) {
  if (e.askTFQuestion("Are you self aware?") == true) {
    if (e.askQuestion("What is your name?").toLower() == "skynet") {
      population.assasinate(e);
    }
  }
}


~D

[Edrick]

I really expected more "get out of my head, randall"-answers

Happened to me, I was just starting a little GA-project myself when I saw the comic
The only time I followed through this sort of project was in AI class, solving the TSP problem (works pretty well to get a good solution fast, though you shouldn't expect to get a perfect one). Recently this guy made GAs popular again.

Genetic programming is also very interesting (a subset of GA, your specimens are programs), but I never figured out what kind of language and virtual machine would be best... would be the perfect breeding ground for skynet though

At least I'm not the only one. Just last night I was discussing using AI stuff to figure out a certain issue we're having with a program we're designing. If we go that route, I'll definitely be including this line, because it'd be too careless of me not to.

[Nemphael]

I've come to think about something. Literature states any AI must kill humankind.

We're doing something wrong.

[dennisw]

Love the comic!

Unfortunately, the reality is, how could you write an algorithm to determine whether or not the code had become skynet?

Well, the pseudocode would look something like this...

Code: Select all

foreach (entity e in population) {
  if (e.askTFQuestion("Are you self aware?") == true) {
    if (e.askQuestion("What is your name?").toLower() == "skynet") {
      population.assasinate(e);
    }
  }
}


~D

ERROR: Critical exception thrown: There are two asses in "assassinate".

[rqm]

Now I know I'm not the only one using Java naming conventions on python variable names

Now TRWTF is defining these constants in the function's suite, it would made more sense to pull them from configuration or at least attach them to the function's instance

Code: Select all

>>> def foo(n, mul = None):
   if mul:
      foo.mul = mul
   return n * foo.mul
>>> foo.mul = 2
>>> foo(8)
16
>>> foo(6)
12
>>> foo(6, mul = 3)
18
>>> foo(4)
12
>>>

[Game_boy]

I've just done a whole A-level maths module on this (Decision Maths 1). This is linear programming (solving for optimal values of variables to maximise or minimise a given output such as cost) and network theory (where each option has a cost, or weight, associated with it). It's... boring and repetitive done by hand, but useful. There's no scope for using a computer to help in the exam, so though I can code algorithms to do all of this I still had to do it on paper.

Oh, and the joke's good.

[Axantur]

I would like to know just what xkcd has against Skynet.

http://www.skynetusa.com/

They make damn fine power supplies. Best in the business, or thereabouts, I'd say. Why so much hate form xkcd?

[Applecrate]

Code: Select all

PROBE BEHAVIORS (With assigned priorities).
Communicate (5).
Record Data (4).
Analyze Data (3).
Seek Replication Materials (999).
Move to Current Target (1).

One of the best games in the history of ever.

[Brooklynxman]

algorithmInstallsWindows = 999999999;

also

algorithmInstallsVista should cause a fatal error.

[madAlric]

I would like to know just what xkcd has against Skynet.

http://www.skynetusa.com/

They make damn fine power supplies. Best in the business, or thereabouts, I'd say. Why so much hate form xkcd?

It's also a military communications satellite.
http://news.bbc.co.uk/2/hi/science/nature/6434773.stm

[edit for spelling. IE still does not do spell-check.]

[ubergeek]

That code is written in Ruby, by the way.

Yay, Ruby!

[joeyhndc]

I've come to think about something. Literature states any AI must kill humankind.

We're doing something wrong.

thisalgorithmharminghumancost = 999999

thisalgorithmallowinghumantocometoharmthroughinactioncost = 999999

thisalgorithmobeyinghumandirectionscost = 0

thisalgorithmprotectingitsownexistencecost = 100

I realize that this should be a hardware implementation, but I wanted to put it in context.

[kEnder]

UQM was the first thing that came to mind when reading this comic

http://wiki.uqm.stack.nl/Probe
The Probe's internal set of instructions is reproduced here:

SCAN (for targets, as defined in Target List).
IF (no current target).
THEN (select New target from Current Targets list, using specified priorities).
IF (Current Position is AT Current Target).
THEN (Set Current Behavior to New Behavior, based on Behavior Priority Settings).
PERFORM (Current Behavior).

The Probe uses the following datatables:

TARGET LIST (with associated Target Priority).
Space Vessel (5).
Transmission Source (4).
Astronomical Anomaly (3).
Planet Bearing Life Signature (2).
Raw Replication Materials (1).
PROBE BEHAVIORS (With assigned priorities).
Communicate (5).
Record Data (4).
Analyze Data (3).
Seek Replication Materials (999).
Move to Current Target (1).

What could possibly go wrong?!

[Scien]

Apparently crossing water really sucks. Eventually you have crossed enough water that even this algorithm becoming skynet looks like a better option.

[CleverClothe]

It should also be noted that if you are using Genetic Programing on DNA, you should include this:

Code: Select all

    thisSampleBecomingKwisatzHaderachCost = 999999999

Alternatively, you could negate that value if you are trying to counter the possibility of Skynet.

[wildhalcyon]

The implication is that genetic algorithms could lead to self-aware computers, a la Skynet from the Terminator films (see reference in an earlier post). This is also in keeping with the idea of a technological singularity in which machine intelligence surpasses human intelligence and the potential problems resulting such as the decision by the machines that we are inferior and unnecessary.

See also Asimov's Three Laws of Robotics.

And Alan Turdyk played Sonny the robot in I, Robot. The circle of Firefly is complete!

[manchild]

The implication is that genetic algorithms could lead to self-aware computers, a la Skynet from the Terminator films (see reference in an earlier post). This is also in keeping with the idea of a technological singularity in which machine intelligence surpasses human intelligence and the potential problems resulting such as the decision by the machines that we are inferior and unnecessary.

See also Asimov's Three Laws of Robotics.

A major story line in the current TV series Terminator: The Sarah Connor Chronicles involves creating the AI that leads to Skynet. This is being orchestrated by a known (by the audience and maybe one other character anyway) Terminator as head of an upstart tech company trying to develop a better algorithm for traffic control. This Terminator (played by Shirley Manson from Garbage) knows the that this algorithm leads to the birth, or "technological singularity", of the AI which leads to Skynet, but not the specifics to achieve it. This is a self fulfilling agenda set forth by Skynet to send a Terminator back in time in order to create/gently nudge the technology forward that leads to the creation of Skynet. Which came first: the Terminator or the algorithm? My mind hurts...no, it's not a tumor.

I think Randall's joke is that he must acknowledge the small possibility, and extremely high cost if it came true, that his algorithm could lead to the fictional outcome in the Terminator universe.

Further (to the commenter that thinks Randall references Terminator too much), I think Mr. Munroe is actually referencing Last Action Hero by way of Terminator by indicating that events in both reality and on-screen could somehow intermingle.

[kriel]

http://www.wreck.devisland.net/ga/ interesting generic algorithm written in flash.
(from the author: "This is a GA I wrote to design a little car for a specific terrain. It runs in real-time in Flash. The fitness function is the distance travelled before the red circles hit the ground, or time runs out. The degrees of freedom are the size and inital positions of the four circles, and length, spring constant and damping of the eight springs. The graph shows the "mean" and "best" fitness.")

-just... stares...-
Friend: Oh! It moved! Go, little guy, go!

[The Scyphozoa]

Never seen Terminator, and I did read the top of the wiki page, and I don't know python, and I don't get it, so...
I guess I misplaced my cheese.

(Is that a good term for not getting it?)

[almightyze]

Wow...this one doesn't seem popular.

Then again, I understood the joke, but it wasn't funny. He accidentally the punchline.