Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[shingekinolinus]

Hi, I'm a programmer and I've been playing with artificial learning a lot recently, and I'm really impressed with how it comes up with really creative solutions. I know this might sound crazy, but I want to apply this to SS to know if there can be some really unorthodox but efficient loadouts.

For those not familiar with machine learning or neural networks, it is basically like this:
- You start off with some random generated creatures (ships)
- In each following iteration, you determine a score for each creature
- Those that perform badly are each replaced by a mutation of a better performing creature
- After enough iterations, you will have a population of well performing creatures

Watch this series if you want to learn more


So in Starsector, you can start off with 100 random variants of a ship. In each iteration, you run a simulation of it fighting against some other ships. The score is determined by how much damage is caused to the enemy ship(s) and how much damage was caused to itself. 50 ships are eliminated, replaced by 50 mutations that have either a turret or hullmod swapped/removed/added. After some iterations, you should have 100 relatively well performing ships against that specific enemy type.

For now, this is just a thought. I'm not a mod developer, so I'm not sure if this is actually feasible. For example, I don't know if battles can be sped up much so that more iterations can be run, or whether the simulation module can be run independently and without graphical rendering.

[Takion Kasukedo]

That's actually a pretty nice idea, but considering everyone who wants to attempt such a thing would have to mod the living daylights out of the game, I thik it should be limited to 30 - 50 maximum, as possibly a quarter or half would be useful.

[Linnis]

Yeah, I guess its do-able.

But because the game replies on AI response instead of natural causes. This will create alot of anomalies in the result data screwing the final result.

I dont mean that AI sucks at using dual light machine gun or anything.

But artifacts between AI interaction will wreck ***, since its not random chance its re-produce-able.


Also there is almost a rock paper scissors thing that could end up making the process take much much longer then expected, and with non-definite results. Its not like... which grip shape is best to pick up a certain shape object.




Interestingly, this happens in alot of competitive games, mostly starcraft comes to mind. But honestly, a well tuned human brain to the task at hand does a much better job in an hour then 20k instance of tries.

[shingekinolinus]

Interestingly, this happens in alot of competitive games, mostly starcraft comes to mind. But honestly, a well tuned human brain to the task at hand does a much better job in an hour then 20k instance of tries.

I kinda agree with you there, it's like I just got a hammer, and everything starts to look like a nail

Also there is almost a rock paper scissors thing that could end up making the process take much much longer then expected, and with non-definite results. Its not like... which grip shape is best to pick up a certain shape object.

To avoid the rock-paper-scissors thing you can run all simulations against the same enemy.

But artifacts between AI interaction will wreck ***, since its not random chance its re-produce-able.

I'm not sure what "artifacts between AI interaction" means. If it means "AI fart", then I'd like to point out that even in simulation there is quite some randomness involved. For example if you run a 1v1 simulation and immediately set to autopilot, the attack angle and enemy response will be different each time, which makes it very hard for the same situation to occur.

[Dark.Revenant]

I would much rather see evolved AI routines than evolved ship variants.  The latter isn't particularly useful, due to the reasons Linnis suggested.

[Linnis]

Well the artifacts are certain instances where AI does really silly things.

A good example is say frigates decide to kite trading blows, they are both steady build and cannot beat each other's shields. Well the ship with more CR would win. Even though, one frigate should have closed the gap and won easy with machine gun fire.

The problem with fighting the same enemy is that your just testing vs one instance. Where a group will evolve to combat one. But the #12 place might wreck #1. But they both have above 90% of beating that one instance.

But if you start cross testing the groups then it essentially is the same thing as fighting random enemies.

[Gothars]

I think the issue with this approach is that its aiming at the wrong end result - AI with great fighting skills. What the game really needs is an AI that provides the most fun for a wide variety of players.

[TaLaR]

I think the issue with this approach is that its aiming at the wrong end result - AI with great fighting skills. What the game really needs is an AI that provides the most fun for a wide variety of players.

I don't think these goals are that different. As soon as you find flaws in AI, combat is all about exploiting them in repetitive fashion.

Plus game has same AI for allies and enemies (which is good). Most infuriating thing is to see allied ship die in situation that was clearly and easily avoidable.