I Made 1.000 A.I Warriors FIGHT... (Deep Reinforcement Learning)

[@Zuzelo]

Who shall I add next: Ranged Warrior or Healer? Or any other class perhaps? Also, don't forget the comment for the algorithm gods!

[@albogd652]

Maybe ranged units will give you more options for different strategies? I feel like healers would just group together and spam each other.

[@Zuzelo]

@@albogd652 hmmmmm the gun goes ratatatata

[@1z34t7]

Ranger

[@donatasjakutis4359]

Rangers will be fast fights. Healers make combat longer if enemy is not killed in first hit. Interesting how will look combat without rangers healers + melee. I just wondering

[@mraxle9028]

idk healers sound harder so do that

[@nikolozgilles]

suggestion! Attacks shouldn't one shot warriors. This will give you more opportunities to punish the ai for being hit and reward the ai for hitting enemies

[@Zuzelo]

In theory it should not make a difference whether the agent is rewarded for killing and punished for dying or rewarded for hitting and punished for getting hit. Also, once I add different types of soldiers, there will be those who do not one shot kill. Since warriors are melee, I thought 1 shot kills make sense

[@irondolphin1559]

wouldn't make a difference

[@itsuko6366]

Maybe give them the option to block for less stamina than a dodge, but it only halves the damage.

[@kuuluna]

​@@Zuzelothat's. Very sad I guess we're not advanced enough for that

[@starhammer5247]

​@@ZuzeloWell, it could be a risk and reward sort of thing, where the AI weighs whether taking one damage to kill one enemy who is low worth it, or it would encourage the AI to retreat.

[@chikkynuggy7522]

You should add a very slight punishment for friends dying, that way they try actual strategies and try to protect each other

[@Zuzelo]

It is worth trying, but I gotta be careful, because it might confuse them too much

[@AttilaDToth]

@@Zuzelo​​⁠ Could do it for friendly death in a radius or the line of sight, maybe some minor reward how well the team is doing as a whole where they are dissuaded from having the team number lower and rewarded for having a higher number then the enemy team.

[@johnwang9914]

​@@AttilaDTothMaybe an affinity score based on cumulative proximity to determine whose death would penalize them more and also proportionally share some of the rewards based on the lowest common denominator of affinities. The affinities should also decay when separated by more than a certain distance. As the affinities accumulate linearly with a given proximity, the weighting of the effects should probably be logarithmic or limited by an asymptote. This could get quite complicated. However, this does foretell the AI apocalypse.

[@Dylan-xv9hx]

**Runs away from teammate so they aren't in line of sight**@@AttilaDToth

[@Birds-are-Cool-j]

@@johnwang9914 I'm just going to leave a little breakdown here for people like me who are a little too dumb to understand this without heavily thinking about it, and don't want to go too deep into thought about it. Although, this was made by ChatGPT, so it could be false, as the only context I gave it was the comment I'm replying to. Affinity Score: Think of it like a measure of how close or connected someone is to others. Cumulative Proximity: The score is based on adding up how close they are to different people over time. Death Impact: If someone with a high affinity score dies, it affects others more. The idea is to figure out whose death would have a bigger impact. Reward Sharing: When good things happen, like rewards, they should be shared among people based on their connection (affinity). More connection, more sharing. Lowest Common Denominator: This means sharing based on the weakest link. So, if someone has a low affinity with others, rewards are shared at that lower level. Affinity Decay: If people are far apart for too long, their connection weakens. So, the closeness factor lessens if they're not near each other. Linear Accumulation: Affinities add up in a straight line with closeness. Logarithmic Weighting: The impact of these affinities isn't always equal. It could be more powerful at the beginning and then less impactful as it goes on. Think of it like a diminishing effect. AI Apocalypse: This complex system could lead to unintended and potentially catastrophic consequences. It's a warning that implementing this might have serious risks. In simpler terms, it's like trying to measure and share how close people are to each other, considering how much it would matter if someone important to them dies or if something good happens. But be careful because this could lead to unexpected and dangerous outcomes, like an AI apocalypse.

[@reiteration6273]

The AI warriors are attacking their allies. Even though they don’t seem to be able to damage them, it would make sense to give them some way of telling apart enemies and allies.

[@Zuzelo]

Technically they are, but I think that is result of the Moshpit training. Since quite often they get spawned right next to enemies. I am curious if they will avoid friendly fire if let to train for huge amount of time (500-1000 hours)

[@arthurorenardeau7231]

@@Zuzelo They will definitely because that just make them lose stamina who is important

[@jasonruff1270]

@@Zuzelo if they see an agent, let them know whether they are friend or foe, if they kill they're own, punish them. I'm currently working on a deep learning project like yours only mine is a bit more in depth (I'm not trying to brag) but instead of swords it's a typical modern army with infantry, tanks, infantry transports and air vehicles etc. I'm working on a very powerful computer with multiple gpus.

[@GodbornNoven]

​@@Zuzeloyes they would but you really need to improve the reward function!

[@johnwang9914]

So until they were punished for living, they didn't seek to fight. That explains a lot...

[@thenerdbird5596]

Just like real life .

[@Zuzelo]

Wait until you see the medic episode :D

[@o1-preview]

Earth Life feature request: Remove punishment for living, increase reward for collective standard of life.

[@FlopgamingOne]

deep

[@senhordoobvio178]

Looks like 90% of wars

[@TrueShadow1982]

"Punished for existing" so accurate

[@Zuzelo]

gotta make it realistic

[@blaizechief2861]

I instantly laughed when I saw the guys fighting against the wall or among allies it made me think of the mindset that AI had adopted, and I believe it's a very hardcore one, like: 'The world is nothing but slander; we must defeat creation itself'

[@o1-preview]

fuck, I've seen similar behavior in some video game AI's i made, during training I started suspecting what you just mentioned.

[@o1-preview]

It started punching the wall and it got real good at punching the wall, when the reward was just not falling over..

[@nonusual]

You should remake this with communication, as in a signals that members of a team can send to other members. This way some signals can help them organize possibly.

[@Zuzelo]

I do like the idea, but I am still trying to figure out a way to do that. Since there are thousands of soldiers, it would be quite difficult to make all of them aware of eachother's actions

[@lucasklokov8728]

​@@Zuzelomake them be able to go through different states at will, and make them be able to see each other's states. A certain state could mean a certain signal. Also it would be cool if there was a general that was aware of everything going on in the battle and could give commands (states) to all of the warriors at once.

[@grandpotato172]

@@Zuzelo give the comunication a maximun range maybe, kind of like if the soldiers where shouting orders at each other

[@bable6314]

@@Zuzelo Make a commander class that can signal, whereas the other classes cannot.

[@thegoatslayer7403]

@@Zuzeloyou should make leaders and commanders to make the forces have more cohesion

[@juanitarodriguez8238]

Kinda symbolic to see them originally not wanting to fight... Metaphoric of real life, and how war only happens due to input by people in power. Maybe AI could lead to world peace someday...

[@Zuzelo]

Indeed, if someone doesnt want to fight for you, just punish them a little more

[@NOSTI2ADAMUS]

It's in our DNA unfortunately... Just look at chimps. They literally go to war against other groups of chimps over territory/resources.

[@Benw8888]

One thing you could do is add a "General" unit that can command nearby allies (I'd implement this by having the General neural network output a vector that is added to a middle hidden layer of every nearby ally's neural network) so that there can be some actual cooperation instead of single agents acting independently. I would like it if you consulted a medieval history/warfare youtuber on what made medieval tactics/strategy worthwhile- then, you can adjust your environment such that those strategies are beneficial. This way the agents learn more complex strategies.

[@Zuzelo]

Ohh comparing it to some real world situations would be POG!

[@Duesikazergaming]

Good work! Didn’t expect your PC could run this and that one soldier boy incident

[@Zuzelo]

Haha, that was an ultimate test for my poor gpu :D

[@argentatidavide]

Great video man! Some suggestions maybe: 1: Halve the damage so that warriors need 2 hit BUT add the back slash hit that cause double damage (so 1HKO again). This should make warriors try to encircle the enemy and on the other hand should allow warriors to stick togheter in order to avoid rear/flank hits. 2. Maybe try to devolop a charge system (Otherwise why a warrior should sprint and consume stamina instead that walking ? ). For example, after a 2 or 3 seconds sprint the warriors does double damage for 1 hit. This should create savage charging and on the other hand stamina management in order to save stamina for a charge

[@Zuzelo]

I am afraid the backstabbing mechanic might be too complex for our dum dums to learn (since they would need to know the relative rotation of all opponents). Charging is efficient for reducing distance (especially when fighting against ranged units), but I think it would be fun to experiment with different sprint modifications as you've suggested

[@rickymort135]

​@@Zuzeloyou could add a sneak attack if they hit an enemy who can't see them they deal double damage this way they should better cover their backs and possibly group up to protect their flanks

[@j.j.maverick9252]

great battle! I think you could include a bit more technical details on how it all works without losing your audience. I’m interested in what options you provided for cooperative tactics?

[@Zuzelo]

The AI can differentiate from enemies and allies, so they can learn that being close to allies leads to more often victories I plan to focus more on cooperation between agents in some of the future ai war videos. Once I add more soldier types, I might make it that soldiers will be able to differentiate not only friends from foes, but also the class of other soldiers.

[@j.j.maverick9252]

@@Zuzelo sounds amazing, I’ll be watching!

[@luckyman4389]

@@Zuzelo looking foward for next vid

[@albogd652]

Oh I could really relate to that punished for existing part lol

[@bmdsch1320]

You earned my subscriber man! Your so underrated! Quality content ! Hopefully you'll upload more of this ai training videos!

[@Zuzelo]

Welcome to the Penguin Gang! :D Oh I definitely will and pretty soon!

[@Schildknaap1921]

Awsome video! I've got one suggestion. Maybe it would help if there was a punishment for inactivity that slowly increases over time and resets to 0 once the warrior kills an enemy. That might make them a bit less passive

[@rickymort135]

He said he added that. They're constantly being punished so need to seek reward fast

[@goldentaleplays9990]

@@rickymort135 That's not what @Schildknaap1921 is talking about.

[@rickymort135]

@@goldentaleplays9990 He's talking about penalizing inactivity, and the reward function already incorporates a penalty for that

[@KamarniWilliams]

​@@rickymort135 goldens right. I think instead of giving them a punishment for always not fighting the enemy, they can have a timer so it causes less confusion among them and less iterations

[@flameingdash]

You should make it so that stats for each character has randomised stats to make things a little more interesting

[@Zuzelo]

I like this! Might be a little more difficult for the agents to learn tho, but is worth trying!

[@rickymort135]

​@@Zuzeloyou could add in a % chance for a hit to register. If the receiver can't see the attacker their % prob of hit goes up. Also make it disproportionately higher if more than one person is attacking you

[@Very_Silly_Individual]

Why do they even have shields if they don't use them???

[@GraysonLovesHisMama]

it just looks cool

[@Shadoh71212]

Looks foine

[@Muichiro_the_fog_hashira]

+Style

[@badcoder6187]

This is exactly what I wanted ❤🎉 You earned yourself a sub like and share. your so underrated

[@Zuzelo]

You da best!

[@animationvreverything3858]

you could add a progressively increasing punishment for getting further from the center to encourage clumping

[@Zuzelo]

I am hesitant at forcing them to do specific strategies, since one of my favorite parts is getting surprised by the strategy they employed :D

[@Kram1032]

I think even minimal motion should cost at least a little bit of energy to make them less jerky in their movements

[@frglcnoly]

I think you should make the same thing of 2 different teams but one team learns from getting punished, and the other learns from getting rewarded...

[@brianmulholland2467]

So....let me get this straight. You're teaching AI ... how to kill? No flaws detected. Perfect plan.

[@nikolozgilles]

the ai be like: LETS GET DOWN TO BUSINESS TO DEFEAT THE HUNS

[@niktniewiem4785]

Damn, that "punishing for simply existing" seems like real life. It gave me existengial thoughts.

[@pinkliongaming8769]

"It already resembles real battles with both sides charging each other" I felt pain when you said that

[@krustygaming5887]

The ai was smart it knew the fighting was a loss and just left.

[@dukeapplecore9401]

I love the accent! I love the content! Sir, you have yourself a new subscriber!

[@Zuzelo]

Thanks my dude! Welcome to the Gang!

[@jasonruff1270]

Any channel that does deep learning stuff I immediately subscribe to.

[@Zuzelo]

Welcome to the gang!

[@speider]

1: When both armies use the same training data, the outcome will either be random or dependent on spawns. 2: A defensive posture with the shield could introduce actual strategy in melee. 5 stamina per second, for instance, with every hit on the shield taking 10-15 hp away from the defender. 3: A quick jab that hurts, but doesn't 1-shot kill seems important. Enough training could teach agents that taking a hit while doing a big swing could be worth it. Or not. More maneuvers, goals, and maybe even scenarios will be needed for this to really become interesting, but I am hoping, so here's a like and a subscription.

[@presidentofthegalaxy3642]

I suggest you add blocking, make it cost 50 stamina to make dodging more effective, but present it as an alternative when stuck in a tricky scenario

[@Zuzelo]

I did :) You gotta check my newer videos

[@mraxle9028]

Okay hear me out i know it's extra work but i think it's going to be cool to show the best soldier of the war like the one who did the most kills like a leaderboard i have a extra idea but honestly idk if it's a good one basically the MVP ai will get a camera to look at him to show the battle more up close and so we can see how well they are doing Oh i think you should make a block mechanic bc they are running with shields and not using them that kinda bothers me a bit Ps truly amazing stuff i honestly rly enjoy your ai stuff it's so intresting and i rly hope that all your hard work will get more appreciation

[@Zuzelo]

Glad you enjoyed it :D Ooooh yeeessss! Leaderboard and cameras for MVP would be definitely cool! I shall try that Haha, yeah, perhaps blocking would have been a better choice than dodging :D In the next Epic AI War video will consider fixing that

[@BloonHero666]

I love how some of them are still managing to unalive themselves by touching the wall in the final battle 😂

[@scaryandscared]

Guy who got double kill deserve a care package

[@santawithaflamethrower2758]

I havent seen your other videos but off the top of my head to improve results, give a minor reward for moving towards the enemy and a minor punishment for moving away from allies. That will encourage proper formation fighting. For now until you can get things looking like a proper battle, keep the same units, but in the future introduce skirmishers, armored units, or archers would be cool

[@Howyoudoing-z2f]

Bro you are Hella underrated you should make this into a game lol

[@Zuzelo]

Haha thanks, making this a game is definitely on the list!

[@JathTech]

You should train the two armies separately first. Split it into 4 armies, train 2 at a time, and after 1000 iterations, switch which armies are fighting each other. This will help to break the symbiotic or parasitic evolution they are going through and allow them to change faster into more refined tactics and strategies.

[@Zuzelo]

What if both sets of armies get to same results/strategies? Since the armies are the same, I expect that to be the case I am using self-play which basically trains only one army at a time. One army trains, the other uses the knowledge, then they switch, so basically what you suggest, but on an unit level

[@Brawl_Legends-ut1vo]

This is so GOOD! You earned a subscriber. 😊

[@Zuzelo]

Welcome to the penguin gang!

[@dashamold1985]

@@Zuzelo i also subbed!

[@chuntingfung6355]

This is very cool stuff. However, is it possible to avoid those in the same team fleeing towards the wall by adding a mechanism for them to detect the closest enemy? That way, they will all charge to the opponents and start a real fight!

[@existenceisrelative]

I like the idea of seeing what happens when individual units have varying degrees of fitness. You'd have the base Stamina of let's say 80, plus or minus 20 decided randomly. Or treat base Damage the same way. Then see if and how it affects smaller groups when fighting. Or maybe give them an objective separate from "don't die" and see how it affects their behavior. Like, if each side had to _reach_ a flag in addition to wiping each other out. A unit existing in proximity to the objective would cause the entire force to gain positive reinforcement. And having an enemy unit existing in proximity to _their_ flag would cause an equal amount of negative reinforcement. So if they just trade places next to each other's flags, they get nothing. Aside from that, i'd want to see what an awareness of their allies status within two squares would do.

[@huraqan3761]

Hey cool project! Performance tip: if you only need to detect other agents you can skip the raycasts entirely. You could use vector dot product comparisons instead.

[@peelipity9073]

Now what i want to see is this but every warrior has different roles similar to Team fortress two

[@gronkaii5057]

Different weapoon types would be cool. Also Different strategies? Maybe train one side with full death match mode 1v1 strategy? This is so cool lol

[@Draktand01]

If you want to make it more realistic, make it so the soldiers have self preservation, but also have an even stronger urge to protect an objective, whether it be territorial control, protecting a commander, civilians, a historical building, or something else.

[@polishscribe674]

Humans: C'mon, the AI won't destroy us, it's just movie fiction. Also humans:

[@biernyvr]

The valhala was so deep dive into human psychology ;)

[@fartninjacats]

the one red warrior in the top left corner is holding up pretty well for a single person

[@happy_turtle1270]

This looks like it would be a fun game

[@Zuzelo]

Hmm, I might just do that :)

[@Y270H]

You could force the wall to shrink (you can either keep the wall lethal on contact, or make it as an actual wall that pushes any one who touches it back...or you can just keep it as a normal wall) the wall keeps shrinking until it has the space for only 1. This hypotheticaly should force more fighting.

[@DanielSMC01-v2m]

The A.I. immediately refused to fight and instead chose peace 🕊️✌️🕊️

[@Zuzelo]

Weak!

[@marcboss]

Great video as always!

[@Zuzelo]

Next one will be even better!

[@marcboss]

​@@Zuzelo If you do an AI battle video again you should add new ways of making strategies or new weapons

[@Zuzelo]

@@marcboss Oh you bet I will. This is a new series on this channel!

[@marcboss]

@@Zuzelo I hope next episode they wouldn't fight air 😅

[@Zuzelo]

@@marcboss Can't promise that :D

[@lololol246]

Underrated!!! The effort was not worth 70k views and 3k likes

[@Zuzelo]

Road to 100k! Let's gooo :D

[@anakintaylor5616]

I would love to see buildings and stuff like that be added that the soldiers can move through or around because it would add a level or strategy to how they think around the buildings, also a ranger type would be very cool to see

[@squirlygeorge]

I would love to see a "capture the flag" element as well. Love the videos though 💥

[@zamknietyK]

Add a flocking behaviour what would cause the solo warriors to have actually chance that they rather want to group up with teammates if he sees them and if theres too muchin one spot then move out of it (if he sees less that 25, then go into this direction every third move, if more than 25, rotate 45-135 degrees and wonder off as usual. It would make it more intresting seeking a little squads from time to time fighting that rather chaotic war. I would also suggest them for making a calls system where the warrior calls out from his range for example 10 meters over warriors to come if he sees 10-25 enemies, if more then to fallback, it would make it really intrestong and the behaviours would basiclaly develop over time. There could be also 5 warriors to occur to be a commander unit what would be basically 2x better than normal unit and it would have 5 bodyguards what actually follow him all the time and attack his target, the commander unit would have also greater fallback and support behaviour distance for example 20 metres. It would be really intresting to see something like that

[@bable6314]

The behavior you're describing is a behavior that will be learned automatically.

[@wendytangli942]

soldier:SIR ALL OUR STRATEGIES FAILED, THEYRE CLOSING IN general:get the frickin computer

[@eeeae]

paused it when the fight began, I bet blue cuz blue is my favorite color and just epic. (Damn it)

[@Zuzelo]

You've tried so hard And got so faaaar

[@vvhitevvabbit6479]

I think you have some issues with your reward function. In order to get them to engage the enemy more often, you should have rewarded the agent for being within close proximity to enemies, and being far away would give a punishment. In this way, the agents would learn to engage the enemy.

[@masterpenguinx5021]

I love these Videos

[@Zuzelo]

Thanks :)

[@prometheusnavagadro4704]

I am not sure what algorithm you are using, but from my personal research, maybe you could implement the COMA algorithm (counterfactual multi agent policy gradients) or LIIR (learning individual intrinsic rewards in a multi agent environment) I have been trying to implement COMA in TensorFlow, and from the paper, it is a pretty good team algorithm to maximize the reward using the actor critic method. Might be cool to look into if your algorithm is taking too long to train, this might be quicker and faster for your agents to learn and cooperate.

[@zyswanson7865]

That was honestly poetic that the blue killed itself

[@MichaelRosenblum_Emp500]

I think something that would have helped would be having a circle in the middle of the map, the closer warriors are to the middle of the map = the more they are rewarded

[@OrangeorApplenapear]

Next time do different classes like bowmen foot troops and maybe even knights. Cool vid love your stuff.

[@Zuzelo]

Knights are on a high priority!

[@critickman]

I was thinking To make it easier for the ai to start getting complex combat strategies you should use the reward as a moral sistem Since what mostly wins wars are the moral and the strategys So you coud have 4 brains Two on each side One is the commander The other one is the soldiers The commander should focus on avoiding caos (example avoiding soldiers to do absolutely nothing during a battle) And the soldiers self preservation and combating it self (adding attack,block) So the first brain makes a strategy (that is the hard part if i find a more easy to understand concept i will respond since strategy is not an easy term to explain a computer ,specially leadership) After the commander brain sets a strategy the other ia should o try to do it It is really complicated now that i think about it So by the time just use the same thing that ants do at a ant war Communications and all

[@Zuzelo]

Having and overseer commander is definitely something I want to try very much, however due to the complexity of the system I need more time to think of a proper way of implementing that. So that will definitely happen, but sometime in near future

[@JarethGT]

Poor last guy saw all his friends die and decided to go out like a samurai

[@varvarith3090]

In the grim future of AI there is only war.

[@Zuzelo]

Gotta start building (and training) my army xD

[@shady2189]

You should add a reward zone.. that disables the punishment for existing.. so that the AI will contest the middle of the map and encourage a fight in the middle.

[@franconullo4228]

I think it would have been far more intresting if all the soldiers were controlled by a singular "hive mind", you could build it by making a model that takes all the inputs from all the soldiers and analizes them with 1 neural network per team, this would enable the model to better coordinate the soldiers

[@ThanRengz]

i havent even watched it yet but by the looks this looks FIRE!!🔥🔥🔥🔥🔥

[@dominicsouthern7672]

Not sure if you gonna do this again but add ranged units like a bow or a musket

[@Zuzelo]

Oh I will and I already did. There are like 5 videos in this series already! Go check them out :)

[@dominicsouthern7672]

@@Zuzelo yeah I wasn't aware but I'd like to see you add some terrain like forests or cities (possibly with enterable buildings) to see how they'd react

[@Chroma-Chrono]

I FR WISH THE HISTORY CHANNEL WAS THIS AND WE WERE WATCHING THIS AT SCHOOL

[@Zuzelo]

damn me too :D

[@thesingularityproject4042]

How does punishing something that can't feel work

[@Zuzelo]

First you teach it to feel, then you teach it to fear >:D

[@lacteye230]

How do you start making stuff like this?

[@Zuzelo]

There have been several attempts before at making an AI War, but I lacked experience at the time (you still can find the videos on my channel), but you just gotta practice. I think the best starting points is looking at the example assets ML Agents are offering and trying to recreate them. If you need to know how to get the tools installed and ready, CodeMonkey has a good video on that

[@allenson321]

judging by how the ai seemed to try to fight its own team, I guessing he did not include a way for the ai to tell friend from foe which resulted in the end a conga line of friendlies following each other trying to kill their allies.

[@Zuzelo]

They are able to differentiate friend from foe, but I made the friendly fire punishment waaay to small I guess

[@12bibbyh2]

to prevent the warriors from running off maybe instead of an instant kill wall you could make the punishment increase the further u get away from the center. encouraging fighting in the middle.

[@bluecollarbionics]

What software are you using to do this do you have tutorials on how to this type simulation ?

[@Combinemetrocop]

This would be a fire game idea.

[@lurkmoar3926]

Yes, I'd like to see many variations on this AI war theme.

[@OfficialDoab]

Man this is cool

[@Zuzelo]

You are cool!

[@5ld734]

Hello I am a rather inexperienced programmer but I am curious why you send 30 rays for sight rather than one ray in the direction of each object in a radius? Are more rays cheaper performance wise than getting direction and nearby objects via overlap sphere or some similar thing

[@Zuzelo]

Having static rays helps the agent learn better, than rays that always shoot in different directions. Moreover, keeping track of hundreds of enemies positions (to shoot rays towards them) is very computationally expensive, compared to shooting a standard amount of rays

[@JKGaming-ec4ni]

did you use a premade learning algorithm or write your own? just curious, im also a coder (unreal engine)

[@Zuzelo]

I use PPO which is the default algorithm offered by Unity ML Agents

[@michalmurgrabia6152]

What about "Capture the base" scenario? Gratification by holding the enemy position.

[@LabdalockVelheimr]

he turned the punishment from losing stamina to taxes.

[@quangan1288]

I think to maximally gear them towards fighting you need to give them a more explicit incentive. For instance, have them try to stay in the center, where the team that has more soldiers in the center gets rewarded and the other team gets punished.

[@leetlebob8203]

You should add an objective, reward the team for staying in the objective and preventing the enemy from getting into it. Or something like that

[@aycoded7840]

You just had to go and make something super cool.

[@Zuzelo]

:)

[@solareiusdeiu]

it requires millions of iterations to make this army intelligent

[@theo1103]

If human players are joined, can it help them learn faster?

[@Zuzelo]

unfortunately you cant train after the game is built :(

[@archanbhattacharjee1122]

Maybe add a moral boost, which increases their attack by a bit for each time they win a fight.

[@MeYou-Sick]

How to make such a simulation?

[@hnl5100]

Next Idea : Try to combine state machine with machine learning in which the AI would have the ability to move to next or previous state ir remain state. However the ai can not interact with some specific states. Alsos a communication feature for the agent would be interesting

[@toz.trendy]

The fact that they chose peace 😂😂

[@KurayTunc]

Great video ! You should make a how to make video my friend looks interesting

[@Crowaltac]

That circling at the end was baffling

[@bruhmomentum6069]

Ok weird idea, what about raising 2 different armies with 2 different natural learning, with the same basic info and combat, but different settings to randomly decide on tactics and test alone -- almost like another nation, which grew seperately.

[@Thy-Chu]

6:14 so far as i know this thing of charge to each other is a mith created by movies, they actually are comanded to especific places to battle, is not like they would just charge, they need to have a formation to it using soldiers with diferent equipements. example: the ones with big shields and spears is in the front the archers a little farther from behind and the ones with swords close to the ones with the shield and spear, this is usually in defense, but the comander who would attack by charging like that is probably not that smart or the place must be really safe because the worst place to enter a enemy base is from the front door. conclusion: just charge is not a good strategy at all, there are some tactics that could be used to make it a better way to attack

[@GeorgeDCowley]

The punishment for low stamina sounds a lot like us.

[@TheNextMarx]

They are like ants in the last formation that’s so cool.

[@050sniper]

did you make any more of these also id suggest punishment for team damage as -25 health for the one doing the damage =) will stop them swinging as wildly maybe ?

[@Philtopy]

I think you may have missed one essential part of a battle: an objective and leaders I think your results would be much better if your AI armies would have areas of the map they really want to occupy and also some soldiers that are more important than others. it would also be a good idea for soldiers to be able to count their allies and enemies in their field of view. This way we will see armies forming formations around important individuals, attacking and defending locations and launching attacks when they have numerical advantage. Also an addition I think of would be to have 3 different unit classes. Fighters (big close combat demage), Shooters (big ranged demage) and defenders (more hp).