Player Features and Ideas Discussion

Since 1943 man tries to simulate brains to learn their MO. That research gave rise to the study of "artificial neural networks"

These days we find neural networks in everyday applications( cryptography , pattern recognition, automation, machine control,...)

Simulated neural networks( software ) can be used in machine learning. It is also extensively used to simulate simple lifeforms, which evolve their neural nets by natural selection, interaction and incentive learning.


My proposal is to give rogue drones simple neural nets, or give the entire hive one neural net.


why?

it will create an ever changing enemy, that evolves over time, which forces capsuleers to adapt tacticts again and again.
it gives researchers a perfect simulated environment to study neural networks, swarm behavior, autopiloting neural nets,...
it fits in the story,


how?

incentive learning: create hunting behaviour : drones need to wreck capsuleers ships for replication**
sensory input: simplified and different from capsuleers:

- drones can see capsuleer ships and drones as objects at a certain distance and angles.
it can "lock" on the objects position and warp on grid.

if a capsuleership is being "locked for warp" The pilot will be notified of a drone getting a lock by some sound in his cockpit

-drones can see other drones too:
if a drone sends a signal trough the solar system ( aka being attacked or found some food ) other drones can react

advantages

- no extensive programming required to create dumb repetitive ai
- actually no programming at all, except the neural net itself: the framework -program running the AI *
- the network learns by the inputs it gets in his environment. In the beginning drones will be dumb but smarten up
- may be a formidable enemy in 3 years ( patching can expand the neural network of offspring)
- please ccp and fellow capsuleers atleast look into neural networks and ask yourself, how could this interact but also what are the dangers*** of neural nets in MMORPGs


sry for the wall of text


http://en.wikipedia.org/wiki/Artificial_neural_network
http://www.youtube.com/watch?v=Rm71Fsp0LCY
http://www.youtube.com/watch?v=g0ZnrzTfFIQ

* edit: you can simply find the classes Neural Network in java, python, mathlab,.., create an object, specify netsize, learning method, input ,outputs , interconnection scheme ,...
and let it run

** our incentive to kill drones will be their loot that ccp can make a key material. fluctuations in drone and capsuleer behaviour will influence price.

*** neural networks allready prove to be very efficient traders, routeplanners, pattern recognizers IRL , and could be used also as exploit . On the other hand: would be cool to use the test server as a server to test neural nets: its a perfect test bed.




TL;DR: simulated neural networks can give rogue drones a very adaptive , intelligent, (lifelike) behaviour.

howmuch bandwidth / cpu would this require from the server/ client?

deleted

Very little, I have a project I want to do that turns an Arduino R3 into a color reader (basically place object and it tells you the color, hopefully I can get it to hex values for the exact shade) and it will need basic training to recognize what full colors are.
But from what I see its very difficult to create but very easy to modify.
Sort of like how creating the computer is a longstanding thing, but changing parts occurs on a 2 year basis

So, 70 years after the business started the only thing which can be said today about artificial intelligence is that it is artificial.

If anything, it says that it is a bit more difficult than plucking neural net sim into game engine and watch things happen. Yeah, pattern recognition works to some degree but behavior is quite another cup of tea. If it was such easy, all shooter games would employ it but we still see NPCs either stupid, scripted or cheating (or combination of any).

Also if it was learning from live game it would be also open to being gamed. Coordinated effort from large group of players done for long enough could teach the net some wrong patterns. While fun to watch for the sake of it, probably it's not what you would desire.

BTW, care to mention where exactly or how neural nets are being used in cryptography?

This is exactly why most PvE players will oppose this to their last breath.

If it evolves and forces adaptation, how are they supposed to farm it over and over again whilst semi-afk?

I don't know. As a partial carebear, I strangely approve of this. It'll bring some respect to a pilot who can continue to adapt beyond the NPCs and even respect (in time) from PvPers observing the strides PvEers take to master fighting their foe.

And it wouldn't have to be that tough/crazy. I'm sure CCP would have limitations on the NPCs (such as reduced ability to vary outside their primary damage type).

neural networks aren't very well suited to produce NPC AI. Black and White had a simplified version of neural networks and you didn'T even notice it. It could have been plain old behavior scripts and it would had worked just fine.

the first far cry had NPCs which changed strategies based on how you played the game. Basically all other games doing now the same since it works best, is maintainable and doesn't have high CPU load.

there are competitions from time to time where bots fight against bots, i don't remember seeing a neural network based bot outperforming classic algorithms - doesn't matter how long you let him fight to prepare the network.

just google robocode and read about how the best bots work, or quake bots etc.

(bot is here used as synonym for NPC AI)

Sorry, as a programmer I'm gonna just state that this isn't feasible -- at all. It just doesn't scale well and CCP would be spending massive resources to make this thing work. Ultimately it's their call, but if they decide not to do something understand it's because of the massive investment it would take without significant results.

No, because Skynet.