Overhaul AI

[NoImageAvailable]

Need to completely overhaul the way combat AI works. Needs to:

• Determine a specific objective based on a predefined raid goal (e.g. kidnap slaves, burn crops, etc.)
• Draft a plan of attack potentially involving splitting up into smaller groups with individual sub-objectives
• Coordinate maneuvers between sub-groups while moving to target
• Determine avenues of approach based on tactical value, i.e. presence of foliage, enemy fortifications, etc.
• React dynamically to enemy maneuvers and reevaluate tactical situation if needed
• Conduct ordered retreat once objective is achieved
• Be able to determine individual pawn targets based on tactical need, e.g. pirate with LMG should prioritize suppressing enemies targeting charging melee and shotgun pirates

[NoImageAvailable]

As referenced in https://github.com/NoImageAvailable/CombatExtended/issues/27 should include AI functionality to control spacing and avoid clustering when possible.

[Valerate]

Not entirely related, but Achtung! has been working on a New combat AI for the your Pawns https://ludeon.com/forums/index.php?topic=22130.0

"A completely novel feature is "Auto Combat". Achtung! has its own build in artificial intelligence that controls your colonists for you. All you have to do is to order a few colonists to their targets and they will get there on their own, keep a safe distance and shoot/position themselves intelligently until the job is done. Perfect for those occasions where you are busy dealing with a bigger issue. For now, this is an experimental feature and with more than a handful colonists or opponents will probably force your computer to its knees. But used smart it will be a useful feature!"

[dvalroth]

a17 is supposed to include new raider AI as well. Dunno if it's worth waiting to see what Tynan does.

[NoImageAvailable]

Just a heads up for newly joined testers: the AI overhaul has already been in progress for a while and large parts are already finished. It simply got delayed because the coder working on it has some RL stuff going on right now.

[kbatbouta]

I'M BACK GUYS!! sorry i've been busy for more than 3 years... I'm already planning on finishing the overall for the AI. (3 years old AI folder was started by me and Nia back in 2017 so it's probably the time to finish what i started )

i'm currently working on:

• Creating a controlled environment in which an evaluation can done and can be reproduced.
• Considering different methods for modeling the desired AI behavior such as
  1. Swarm modeling.
  2. Reinforcement modeling.
  3. Adversarial modeling.
• Finding suitable "Matrix multiplication library" similar to NumPy for C#.
• Defining a differentiable mathematical function for the AI "Strategic goals".
• Defining a differentiable mathematical function for the AI "tactical goals".
• Defining a suitable feature-space for the optimization problem.

[NoImageAvailable]

@kbatbouta are you that AI coder from Syria from way back? I thought you died from that stroke

[kbatbouta]

@NoImageAvailable yeah, it's cancer. So is it possible to automatically spawn pawns ( or at least to do so in large numbers automatically without having to continuously adding a hedif for each ).

Our best bet is to use evolution for creating a toil sequence. Such as we let a decision tree a create number of random sequence of toils and have them compete with each other. Then using the top 1% of these to retrain the decision tree. Repeating this process for a large number of times ( many generations with many of the older one competing against the younger ones ) would yeild us a model capable of creating the "AI" for us.

also the normal AI can also compete with them. basically a battle royale for AI to create other AI.

for determining the final difficulty. random noise could be added to the decision weights, the larger the noise the less difficult it is.

also for what targets to choose and what environmental factors to take into account as of now just the "stock order of operations" while only modifying the order of actions.

this data would show what factors must be added for considering. using a Scikit-learn type of library is preferred for now since most of the first stage is the (i'm going to call it this for the sake of simplicity) battle royale implementation phase.

sources:

• https://en.m.wikipedia.org/wiki/Decision_tree_learning
• https://en.m.wikipedia.org/wiki/Negamax

On Wed, 11 Mar 2020, 5:54 am NoImageAvailable, notifications@github.com wrote:

@kbatbouta https://github.com/kbatbouta are you that AI coder from Syria from way back? I thought you died from that stroke

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/CombatExtendedRWMod/CombatExtended/issues/73#issuecomment-597615047, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABPTV7HFFXLAP6IZARCJE5LRG6CYPANCNFSM4DB6DKJA .

[NoImageAvailable]

Talk about bad to worse. Anyways, I'm not actually involved with CE anymore since I'm working on my own game now, but if you're still interested you might want to pop into the Discord and talk it over with the current maintainers.

[kbatbouta]

@NoImageAvailable Oh so all the great elders and veterans have left rimworld. yeah i will, just need to find my discord account. Also you can't mention a game with no trailer... and no preorder url cuz by what i remember of your standards: take my money or i will force you.

yeah do you still have the discord server name?

[NoImageAvailable]

CE Discord: https://discord.gg/safyWuA

As to the game, it's still in a pretty early stage. Did a lot of prototyping, now porting to Unity's ECS for multithreading/efficient memory management for an Early Access release hopefully starting April. Got a website and Discord here: https://www.vivid-storm.com