Generation Generals

Let's start with a neural network. It can have back propagation to allow it to learn a little bit while executing, but overall, we will train it with a genetic algorithm. The input should be data representing where its units are on the field, what intelligence it has, some data on the environment.

Once we have such a network, it is time to train it. For the first training round, the network, just keep working at it until it can move its units around without having them kill each other. It may be hard to assign a fitness to a particular network configuration when they all seem to move and shoot randomly, but there will be some that quantitatively perform better than others, so find a way. Once the network has some coherence, have it try to kill a single enemy unit that does nothing, but be sure to have it randomly placed in the field so that the network doesn't learn to just shoot in one direction.

The next challenge should be a group of enemies that behave like zombies. They just lurch around the field not doing much besides getting shot at. Fitness can be awarded based on how quickly the enemies were all eliminated, how much energy was spent killing them, and how many resources were used to do it. Once acceptable at that, the enemy should be slowly upgraded to have better tactics. They should hide behind obstacles in the environment and run from the network's units when attacked.

Now add some more features to it. Give it the ability to have simple simulated battles against simpleton opponents, such as nigger street gangs with poor offensive capability. It should have superior numbers and capabilities. Gradually decrease its advantage and increase the strength of the opponent until the network can dispatch a few hundred enemies with few units of its own. Introduce variations of this challenge. For instance, give some units sticks and rocks to fight with, while others have machine guns. Keep the battle varied so that the network doesn't learn to depend on any ability too much.

Once the network can defeat an army of lots of dumb targets with only a few units, increase the enemy's intelligence. They should no longer blindly run around while hiding behind trash cans, they should divide into squadrons to attack from different sides. Give the network the advantage until it begins winning, and then begin to decrease it. Once they are defeated with ease, upgrade the enemy again.

Soon, the network should be a passible tactician. It is time to start real training. Have it fight opponents that would act as real military leaders do. Have it fight through history's famous battles with a variety of capabilities and in all environments. Have it fight Caesar, Napoleon, Hannibal, Alexander, Genghis Khan, Lee, Attila, Fredrick the Great. Introduce supply lines, weather, fatigue, and espionage. Run many simulations in parallel and make the network fight against a slightly mutated copy of itself. Have the top contenders replay hundreds of large scale wars thousands of times, each time getting better.

Now teach the network of asymmetric warfare. Give it just a few hundred men, each willing to die, and have it fight a much larger enemy. Train it with the tactics of Al Qaeda, the French Resistance, the American Revolution, and the Viet Cong. Assassinations of key enemy units and war through attrition are now possible. Have the network play both sides. When it has gotten a grasp of how to fight a much more powerful enemy effectively and possibly win through attrition, cut off its resources and make it earn its keep. Train it to act like a drug cartel or a mafia, where it has to generate its own sources of income to continue the fight. Keep training until it has a good command of economics.

Next, introduce disloyalty, morale, misinformation, and an unaffiliated population to the mix. Subtract from the fitness of the network when it commits atrocities. Units should sometimes defect when the population's approval is too low, or desert when faced with a superior enemy. This should apply even when the network executes a bombing campaign based incorrect intelligence claiming that an enemy munitions cache is located at point X when it is actually a civilian hospital. The network should have to hold conquered territories when in a large war campaign while advancing, possibly facing resistance from the local populous.

Now it is time to show the network the horrors of war. Insanely powerful weapons and evil enemies (based on previous iterations of the network with most disadvantages removed) should be arrayed against it. Enemies should be armed with nuclear and biological weapons. They should run scorched earth campaigns against both captured territory and their own. They should strap children onto their trucks to dishearten the network's units. They should poison water supplies. They should massacre every person in captured territory. The goal of the network on this level is to defeat the enemy while minimizing casualties and destruction, as well as positioning it's armies and navies to withstand powerful strikes, such as nuclear attacks.

By now, the neural network general is nearing completion. Have it fight against itself with all restriction combinations and with many different configurations of its forces, again in a genetic algorithm fashion where the losers are destroyed and the winners are copied. Be sure to prevent inbreeding by having a large pool of divergent network bloodlines that fight against each other.

Slowly introduce more features to be exploited and scenarios of attack for them to learn. Try to make the simulation more representative of reality with each passing iteration. Perhaps you can upgrade the general from only warfare to the ruler of a nation, in control of the economy and the workforce. Or maybe you can have the network fight against an alien invasion that disrupts much of its pre-battle preparations. You might place it in control of a private military corporation where economics is directly tied too military might.

The endgame here is whatever you want it to be. If you like, you could build a robot army and have your network command it to achieve the goals that you set to take over the world. Another interesting idea is to integrate this network into a larger network and use the predatory and tactical skills in some other context. They sky is the limit, and there are few AI applications that having a general strategizing module would not benefit.

It needs a heart.

It needs an anus.

It needs a hug.

>>4
*hugs*

OP doesn't even understand how a transistor works.
shit thread

Ender's Game for a computer.

>>1

Let's start with a neural network

Neural networks are shit.

http://www.i-programmer.info/news/105-artificial-intelligence/7352-the-flaw-lurking-in-every-deep-neural-net.html

>>8
that's not really unexpected is it, for 'Offline' learning?

Neural networks are a made up "computer science" concept for a style of organization of computer data and not in any way related to the neuron system of brains. That article contains shit about java and python so obviously it isn't hardcore.

>>10
"Neural networks" is just a marketing term for computer science, just like "putreficial intelligence" was in the times of LITHP and just like "cunt 'em computing" is now.

>>8
Then why is Google using them to make programmers obsolete? Enjoy your final days, meatbag, NN are the future.

http://www.newscientist.com/article/mg22429932.200-computer-with-humanlike-learning-will-program-itself.html

Other researchers at Google are also trying to teach computers to learn more complex processes. One team recently published details of a neural network that is capable of learning to read simple code and execute it without first being taught the necessary programming language

>>12
lol google is a stupid nigger corp and "simple code" pop sci rofl

>>13
Who should I trust here? The large corporation with thousands of experts working on it in a plot for world domjnation, or a French troll who rarely ever gets more than a pity response.

>>14

trying to teach
learning to read simple code

I'd say the large corporation and the French troll don't really disagree with each other here.
Neural networks have been a buzzword for decades but the results are still largely inconclusive.

>>15
That article is shit, read the papers. It's both less impressive than they claim, and more important that they make it out to be. NN may have been a buzz word, but it was not one that very many people took seriously since the AI winter, when they were little more than cellular automata or fancy Markov chains with a little graph theory thrown in. Their failure was caused by circlejerking academics who were more interested in having ethics meetibgs and writing grand proposals to the DOD than studying the material. It's good that the old ideas are being reevaluated, even if this doesn't work out. Hopefully, Google has enough sense to tell the social scientists when to shut the fuck up so that some actual programming can be done.

Furthermore, it is not nessacary for NN to be exactly as a brain neuron would work, so long as they are turing complete. They could model a brain at the quantum level and still get results, albeit slowly, if it is.

Background on Neural Turing Machines and their performance compared to regular neural networks. It can memorize sequences and replay them almost perfectly, as well as sort data faster than quicksort.
http://arxiv.org/abs/1410.5401

In this one, the describe using a different NN setup to predict the output of a segment FIOC.
http://arxiv.org/abs/1410.4615

>>16

They could model a brain at the quantum level

wat
go back to your transhumanist singularity pseudo-science community.

>>17
Anything Turing complete can simulate any else that is Turing complete. You should know this, you dumb twat. Kill yourself.

>>18
Which means I should be able to simulate a quantum one trillion transistor neural network in my head, nigger.

implying such a computer wouldn't still be anything other than a niggerton hunk of mindless silicon

where is your computational and digital god now, nigger

>>19
You could simulate a quantum one trillion transistor neural network in your head, nigger. You would just require a humongous amount of time. The Turing Machine is a time-agnostic abstraction, it doesn't give a fuck about your mundane problems like getting old or stars blowing up into an explosion of throbbing cocks.

>>20
You could not, as human brain has very shitty and unreliable memory.

>>21
You could use a really long roll of toilet paper to write it all down. The Turing Machine is a memory-agnostic abstraction.

>>18
this has nothing to do with turing machines

it's pathetic that understanding babies first theorem in CS makes you feel this powerful and superior to everyone, who, by the way already understood this decades ago.

>>21

I cannot, as my nigger brain is very shitty and unreliable nigger memory.*