orbit-wars

[dry forum]

hey oh! a question, how every often do the agents fight? I had my first submission that it did quiet a few in a short time but i had another submission doing less in a way longer time so i'm confused now , there are more participants but i had less fights in a longer period of time

[ripe patrol]

Hi! Can I we plz have code of the game?
It is really useful to optimize agents

https://github.com/Kaggle/kaggle-environments/tree/master/kaggle_environments/envs/orbit_wars

Happy to say, I have an answer for my own question

[unique niche]

Hello,
I'm trying to use java (like in lux competitions). But my submissions fail.
I've created a discussion https://www.kaggle.com/competitions/orbit-wars/discussion/692898
If someone has used java or another language (like in lux) sucessfully, it would help a lot.
Thanks.

[ripe sedge]

Hi all

[ripe sedge]

Has anybody summitted any agents yet ?

[spring owl]

Yes

What I'm quite worried about is that my agents are not performing up to standard

Like for the same agent I submit now I got 200 points less than the agent I submitted previously

[ripe sedge]

I got my first agent submitted 🙂 I'm assuming they make it play itself as a validation ?

[spring owl]

Yes

[storm wadi]

Doesn't the score depend on the field strength? So it would be expected for the same agent to show lower score as others improve

[spring owl]

That's true but I'm really running out of ideas on how to improve my score 😞😞

[lucid orbit]

Hii @everyone 🤝

Myself Amber Agey from India

[glossy schooner]

Hello I’m seeking teammates dm me if you’re interested

[opal spear]

Hello all!

I'm wondering if someone can answer a question for me. How might you submit a bot that uses external libraries, e.g. numpy for this competition?

[ripe sedge]

I woul just test it Im sure numpyt would be available in the enviornment

[opal spear]

Makes sense! I just realized the kaggle_environments install includes a ton of stuff, numpy included.

[drowsy ferry]

hello. im happy to be the part of this competition. i have a question. what version of orbit wars that used in orbit_wars competition right now?

[spring owl]

Sorry what do u mean?

The rankings will keep on changing based on the agents launched by the participants

[drowsy ferry]

sory ,i mean orbit wars version. becuase now im using orbit wars 1.09

[spring owl]

I don't think there are versions??

[abstract junco]

Yes, the leaderboard is running version 1.0.9. There should be an update with fixes for 4p incoming on Monday

[viscid canopy]

Has there been more 4P games recently?

[drowsy ferry]

owh, im lucky. here i use pydroid to run the simulation

[sand orbit]

did the 2500 rated guy get banned??

💀

[pulsar barn]

no, hes on 4th place now

anyone have any luck training rl agents?

[river hedge]

I've tried A2C before, didn't fully know what I was doing though

Now I'm experimenting with evolutionary algorithms because they seem more intuitive to implement (at least for me). It's really slow, but I got a score of 700~800 with 500 generations and just self play

[pulsar barn]

is the agent still a policy network, or is it a hardcoded python agent?

im trying to train an agent with ppo, but its really slow to converge

[river hedge]

It's a two-layer network implemented with numpy (because GPU acceleration will not help, as the main bottleneck is the environment, as far as I know)

What helped me was better selecting the features for the model (fewer and informative features make the model output faster, avoiding timeouts and speeding the training)

So my code looks like this:

1. Make a list of N discrete possible fractions of ships to send
2. Make K tuples of possible combinations of (source, destination, number of ships) as possible actions
3. Calculate features based on each action and the current state
4. NN should return probabilities of all actions given the current state
5. Take the most probable actions
6. For each action, calculate the angle between source and destiny numerically (I don't need the network to learn this parameter as it would overcomplicate the approach)

[pulsar barn]

i can understand using evolutinoary algos for architecture discovery, but you still have to train the net, right? like are you using the algo to optimize the net?

[river hedge]

There is no optimizer and it's gradient free

Play agent vs agent → calculate the fitness of each agent → cross over best agents (combine vectors) → mutate (add noise to the weight vectors) → repeat

[pulsar barn]

sounds strange to me, but if it works time will show

[river hedge]

Yeah, I don't think it will win, but it is fun

And it's generating good results so far

[pulsar barn]

cool cool

i think RL is the way to go though. just gotta tune the hyper params, make the env easier to learn etc. then just let ppo converge at optimal policy

[river hedge]

Yes, probably. Good luck!!

[pulsar barn]

Good luck to you too!

[misty oyster]

looks like lots of folks are rewriting the env in jax or cuda (could help the env not be the bottleneck)

[pulsar barn]

what kind of reward shaping are people using here?

[hollow hollow]

Main reward is the win/lose one,and i adopted some aux rewards like ship difference, fleet hit reward.(But i scaled it in order to make 0-sum) As far as i guess, aux reward will not be important after the model become strong enough. until then, aux rewards would help to improve faster, but after that, it would prevent learning strong strategy

[river hedge]

I will certainly try, thanks for the suggestion

[pulsar barn]

I figure reward shaping is essential for early learning, it shouldn't have to prevent learning a strong policy, just put more emphasis on terminal rewards

[earnest birch]

how do i start implementing RL? im being stopped hard at ~1050 right now using rule based

[pulsar barn]

Thanks for the tip, im getting 6-8x speedup with a rust implementation of the env. This will really make RL feasible.

[queen glade]

Hello guys

[queen glade]

Hello,
Have you seen any resources for building RL

[pulsar barn]

use a framework like stable-baselines3 or similar, lots of resources in the docs also

the youtube series Foundations of Deep RL with among Pieter Abbeel etc is a good place to learn the basics

[ember sinew]

hot take: I think it's imprudent to attempt ML/RL without having a heuristics bot in the top 100

[pulsar barn]

I think it's imprudent to even make a heuristic bot for a kaggle challenge.

[earnest birch]

i havent 😭 ive paused for now since finals are coming up but i will resume once im done

[earnest birch]

i kinda agree since having a heuristics bot in the top 100 give you a pretty good understanding of how the game works and how to win in it

and that could translate over to ML/RL since you already know what to do, training it will be a little easier (though i cant say much about it since i havent tried it yet)

[queen glade]

Okay

[pulsar barn]

There really isnt much to understand about how the game works. I'm here to learn ML, AI and datascience, not python programming. I guess everyone has different priorities...

[ember sinew]

I think a prerequisite to building strong ML models is developing familiarity with the domain. I think writing heuristic bots at the beginning is the fastest way to attain that. I'll probs explore RL a bit later. Best of luck!

[pulsar barn]

I get what you are saying, but having a bot at top 100, heuristic or not, is more than just exploratory. You could be sinking too much time into it, that could otherwise go to train an RL bot. Best of luck to you as well!

After a week of tuning and testing I finally have an RL bot that is actually learning. Still will take days to train a competent bot with my setup

[ember sinew]

Learning better abstractions / transformations of the game that let you build better feature + action spaces seem critical. Can make the learning task half as hard if you have a better feel for the game

[ember sinew]

how do you handle the action space?

[pulsar barn]

multi discreet masked action space

[ember sinew]

so each edge, 3 one-hots for 0/needed+1/all? and do you combine that with any search or just single inference -> play move

[pulsar barn]

launch, edge, ship fraction bin

i search for the best angle, too hard to learn the angle itself

my feature extractor builds candidate edges that policy chooses from, togheter will a launch flag and a ship fraction from a bin. this way i can use MaskedPPO from sb3. then i search the angle and convert to native action

[pulsar barn]

what is the strongest open source heuristic bot right now? want to do some benchmarking

[thorny mason]

probably orbit wars or something similar - you are not doing some form of self play and heuristic in compedium? i keep hitting a plateau and dont get anywhere beyond 10% v my heuristic

[flat sage]

Are you guys going for a 100% win rate against random bots before submitting for live gameplay, or just winging it?

[misty oyster]

You get 5 submissions per day, so depends on how miuch you are planning to work on it in a specific day. The leaderboard is a very good feedback mechanism, watching your games there can often show you what to work on next.

[flat sage]

Good point!

[pulsar barn]

Im doing a curriculum. Just going hardest from the beginning is not optimal for speed. Self play kicks in at later stages

[thorny mason]

@pulsar barn how are you building your curriculum i have been trying with different ways - i wanted to try and do steps 0-50, 0-100, 0-200, 0-500 because its so sparse and i was hoping to get to like 70% v a good heuristic in each step before progressing but no chance 🙁

• so i used ppo initially then added awr then curriculum - but i am just not getting above 65% or so on the self play - using the baseline and a lagged player he should beat and the heuristic not more than say 15% for the first set of steps

• i am using a delta of fleet + prod * potential prod of mine - the opponent :- really not getting far

• am more than open to ideas as i am a little stuck for how i can improve - there isnt really enough data i think hence my idea with self play as much as possible

[pulsar barn]

First i benchmarked all the heuristic agents i have gathered, then i introduce them by hardness slowly throughout the learning run. i also increase four player probability over time. i dont go to new stages before i reach a certain win rate

i also do pretraining with imitation learning to bootstrap early learning, saves me many steps

hope by next week i have a competent agent. self play is introduced much later because it kills my throughput significantly

ppo is hard to do because it only works within a sane range. i therefore made a pid controller to adjust learning rate to keep approximate kl divergence within a range

[thorny mason]

yeah i am watching the entropy and kl - any luck so far - i mean i am assuming the reward function is gonna be the killer, i assume you are using some sort of dense function to begin with, i tried vs random and a few others and it kinda learns to optimise fleet and planets but against the heuristics they are just too good - you are probably right that i will need to introduce a load of not so good agents for it to learn to beat progressively

[pulsar barn]

i use a dense reward yes. score is dense reward signal so you dont have to only do win/lose. but i cut back shaping trhoughout the curriculum

[thorny mason]

ok - sounds interesting 🙂 i may give your pid controller and the weaker to stronger opponents ago from my baseline - do you have any other tips? 🙂

[pulsar barn]

i think the main thing is the curriculum. because of the sample inefficiency of rl, trying to help the model as much as possible to have a smooth learning is crucial, unless you have unlimited compute.

also, rewrite the env in a high performance language like rust. it will really speed up throughput

[thorny mason]

it has dramatically improved! i have passed through plenty of gates already now onto the 200 steps against where i need to be beating my better heuristics! havent tried it on 4 player yet tho...but i have 2 very weak and 2 weak and they are being destroyed!

[pulsar barn]

nice man!

[thorny mason]

thanks for the input!
--- Eval checkpoint (ep 500, horizon=200) ---
[tier 1] vs bully: 4/4
[tier 1] vs prospector: 1/4
[tier 2] vs rage: 4/4
[tier 2] vs dual: 3/4
[tier 3] vs baseline: 0/4
[tier 4] vs shunlite: 0/4
[tier 4] vs v131_2p: 0/4
[tier 5] vs v131_denial: 0/4
[tier 5] vs v131_wave: 0/4
Eval total: 12/36 (33.3%) improving 🙂

[pulsar barn]

may i ask what framework for rl you are using? or is it a custom implementation?

[thorny mason]

custom

[pulsar barn]

@thorny mason you can find some more opponents here: https://github.com/automatylicza/orbit-wars-lab/tree/main/agents/external

not my github, but you find it in the kaglle discussion board

[thorny mason]

cool

this is where i got some from the old planet wars winner https://github.com/melisgl/planet-wars/tree/master/example_bots

they are in java and not exactly orbit wars but llm of choice can port them 🙂

he also has a great write up of his tactics!

[pulsar barn]

ill check it out

how many steps per secound in the env are you getting total?

[thorny mason]

i am getting about a game per second so currently 200 steps

and i am running a 2p and 4p training harness at the same time

4p sucks tho looks like it is learning nothing

[pulsar barn]

i sample 2p or 4p games at an increasing rate through the curriculum, so during a rollout i will get some 2p and some 4p based on a percentage

[thorny mason]

ah with just one game so 1 model that will play both? from my heuristic i have also split the logic because i find the tactics are so different - i figured also for the rl

[pulsar barn]

yes, my model plays both types of games. starting really low with only 5% 4p games, up to 50% which i believe is the true objective with reagards to 2p vs 4p

[thorny mason]

interesting...

[summer glacier]

it is in lisp, not java.

[thorny mason]

the agents in java i believe

there is some lisp - but same same for a llm 🙂

his write up is gold tho

[grizzled junco]

@everyone so how do you train models to play this game

Like my basic strategy is make a function which calculates a distance from my planet to a target, and based on the number of turns, targets either rotational or static planets, high production, only settling for low production if all has been taken

And if I send ships, I send no more than 20 percent of the planets total quota

[modest bear]

does production change over time, for planets?

[pulsar barn]

i train it with ppo implementation from stable-baselines3

[pulsar barn]

i think it is constant

[soft sage]

Yes constant

[grizzled junco]

But where'd you get the data though

Like first you train the thing and then what? Like I coded a function the environment runs but i dont know what next

[pulsar barn]

PPO is reinforcement learning

you get the data from training...

every rollout generates data

[loud lynx]

the bot plays against itself

[short mist]

guys is there any really good bots i can download to use as opponents for tests?

[pulsar barn]

enders fleet is pretty good, also marco dg v3.3

those are probably the best bots in the forums right now

[humble cloak]

Hi,
I wanted to ask something.
Each fleet is represented as [id, owner, x, y, angle, from_planet_id, ships].
If the owner here is enemy, is the angle leaked, or is it a placeholder only?

[cosmic hatch]

It's not leaked, rather deliberately present to help us calculate the position of ship and where it's headed to

Further, It can help you find out the nearby enemy planets if you verify the coordinates and trajectory (mostly ships go directly towards the enemy)

[humble cloak]

But if that's provided for the enemy fleets, then shouldn't it be called leaked? Because I dont feel there's game anymore. Everyhting is visible.

[cosmic hatch]

Not leaked, think of it as modern day warfare.. Country A and B both have high tech machinery and as soon as Country B sends a missile towards A, the defence system of A detects the angle, coordinates, point of origin and what not to narrow down the threat

It's not leaked, but calculated by the defence (radar detector)

Your focus should be on to narrowing down the threat, the real game tbh.. threat detection is a work for radars now

[humble cloak]

Thanks for that POV.

[cosmic hatch]

Yup 🐣, welcome :)

[pulsar barn]

Would you call Chess a game? It's perfect information too...

[grizzled junco]

So when we figure out where to send ships, dont we multiply the angular velocity of the target planet with the speed of my planet ships in that pair or s o mething?? Then we use the radius and distance from sun and we do some polar type crap to find the new cartesian coordinates and then we compute the distance??

[cosmic hatch]

Umm, I don't quite understand what you're trying to convey or the formula you're describing

From my knowledge of kinematics and circular motion, I don't remember any formula where we multiply "angular velocity of thing with speed of any thing" to get a proper pair

I don't know about this formula, but upon doing dimensional analysis
Angular velocity(radian/second) * speed(metre/second) will give (m/s²) which is acceleration (i guess centripetal one) and upon integrating acceleration twice with respect to time we can get to position. But still I'm very unsure about your formula, can you please elaborate it or was it just an assumption?

And about finding new Cartesian co-ord, are you referring to (x = Rcos(theta' + omega.t), y = Rsin(theta' + omega.t))?

But anyways , here's my mathematical intuition:

First we determine the start(q) and destination(d) of ships:
Assume ship starts at: (x', y') and target is at (x'', y')

For any circular motion(here planet) we can say
x = x(original) + vt
or, in terms of circular motion:
Theta(final) = Theta(original) + omega*time
Here, theta = angular position & omega = angular velocity

And then the Cartesian co-ord would be:

 x'' = R.cos(theta(original) + omega*t) [horizontal component]
 y'' = R.sin(theta(original) + omega*t) [vertical component]```

And we know that `distance (d) = speed.time`

```distance with target t ≤ d```

i.e.

```underRoot[(x'' - x')² + (y'' - y')²] ≤ s.t```

This should give a minimum `t` i guess.

**Note that I still haven't properly tested this out, on large scale data, but it did work for a few tests. I might not be correct 100%)**

[grizzled junco]

New angle = initial angle + angular velocity × ship speed

Then x = r cos new angle

[cosmic hatch]

Oh that's the same formula i mentioned, just in terms of theta(angle) and omega(veloctiy)

[grizzled junco]

Y = r sin angle or something

[cosmic hatch]

I get it, we have the same explanation

[grizzled junco]

But i dont know... there was this demo with this generic function returning moves where you just input into the environment variable

What is with these people training models and how do I implement it? 🤔

[cosmic hatch]

This one (orbit wars) is an ongoing competition, all models available for training are demos and mostly we're expected to do something better to get better results and win

I don't think we can use someone else's model for this 🤔

[grizzled junco]

So you make ur own model woth pytorch or something?

[cosmic hatch]

It's like, they're giving us a bunch of data (a dataset basically) and values as mentioned in Komil's message #orbit-wars message

And then we're expected to do the math and build code to reach the highest score

Every win increases the score, and every defeat decreases it...

And yes we're training a model but not with PyTorch, we're making it accurate with maths

Check more here: https://www.kaggle.com/competitions/orbit-wars/overview

[cosmic hatch]

Ugh, I was finally able to attain a good score.

But how do I coordinate simultaneous attacks, anyone got any suggestions ;-;?

[grizzled junco]

Could I take a look at ur notebook?

[grizzled junco]

Group instances of a compute moves function or so.ething and have them return moves

Like with parallel processing or async?

[cosmic hatch]

I'll share it soon, lemme just edit some more stuff

[cosmic hatch]

Parallel processing might slow it down I doubt, I found something else, I'll have multiple planets attack one target that is closest then decide if they should attack same target or upcoming threats can be handled easily and individually

Basically pseudo group work

[queen glade]

If you were to build a reinforcement learning model for this competition how would you do it pls

[queen glade]

Cause it seems like some pple are training RL models on this task

[pulsar barn]

id start by writing the a gym wrapper for the kaggle env, then pick a rl method from a framework and run it

[queen glade]

Ok will definitely try this
👍
Thanks man

[raven vigil]

Anyone struggling in Consistency and want to learn together.
DM me.

[bold harbor]

hi

i am participating today, is it too late?

[proper crystal]

nah there's like a month left

also, I remember seeing that someone made a byte-accurate faster simulator for the kaggle environment
I wonder if anyone has the code for it

[brazen swallow]

It shouldn't be too hard to recreate the simulation in c/cpp then bind it to python

[cosmic hatch]

i got someone's feedback in the same channel, they refactored it with rust and it was 6x faster; I followed the same, was just 4-5x faster for me but its worth a try

[proper crystal]

I see
thanks ya'll

I might just write it in JAX, vectorizing what I can, and call it a day lol
hopefully that works

probably won't be exact but whatever

[cosmic hatch]

jax cool lol

[proper crystal]

yeah
I'd be able to choose my accelerator without bothering with reoptimization

or torch.compile

[cosmic hatch]

i understand that

it'll be helpful

but im trying to find more ways to improve the performance

[proper crystal]

collision checking on GPU might be significantly faster than on CPU

especially with RL environments when you're running a bunch of sims at once

I'm currently just building a JAX wrapper for the provided env and probably will work on a JAX sim reimplimentation

[cosmic hatch]

but GPU has usage limits : (

[proper crystal]

I vibe coded something a while ago and it had a 10^3 throughput speedup
but it made some approximations that I didn't like
I'm gonna do it myself later

[proper crystal]

I'm kinda just using colab and my local GPU for now

[cosmic hatch]

10^3 damnn

[proper crystal]

and will transfer to Kaggle later

yeah it was crazy

[cosmic hatch]

thats some crazy vibe coding

what ai agent was used?

[proper crystal]

GPT 5.3 Codex via Zed

but I think it didn't model comets properly
among other things

so a realistic speedup from a GPU implementation of the sim might be less good

but even a 100x speedup is insane

[proper crystal]

it also wrote a multiprocessing + numba implementation that achieved a 10x speedup over the base env

it wrote a byte parity test and according to that the numba implementation was byte accurate, but idk how much I trust it

the numbers seem to good to be true

but I still believe that there is something of substance that can be achieved via GPU acceleration
even if numerical precision may be off

[cosmic hatch]

oh i see

[cosmic hatch]

try to prompt hard code, feed it some examples about the comets yourself

[cosmic hatch]

yea that was so cool

i dont like vibecoding, but tbh its cool at times if used properly

[proper crystal]

It wrote comet code for literally everything but the JAX implementation

It kept claiming that comets weren’t easily modeled on GPUs

I found that LLMs struggle with JAX a lot

most of the time I end up just doing it myself lol

[proper crystal]

yeah, me neither, was just feeling lazy
I’m not satisfied with the end result

I cannot get coding agents to think in vectors and math for logic no matter what i try

[cosmic hatch]

try with the rust one next time, and if it persists its claim then its probably on limits

[proper crystal]

wdym

[cosmic hatch]

like refactor the env in rust

[proper crystal]

oh I see

[cosmic hatch]

yupp

[proper crystal]

I mean I know it’s claim is bs because I know exactly how to model comets along with planets

[cosmic hatch]

that happens to the best of us ;-; we're brave

[proper crystal]

it’s just that I was feeling lazy

it claimed that a bunch of things weren’t vectorizable until I told it how

[cosmic hatch]

they have limits, might have to feed it the winner's solution to get it to work lol

[proper crystal]

lmao fair enough

might try it
but honestly I’d rather have code that I can understand atp

so I can write better tests and such

[cosmic hatch]

tru that

[cosmic hatch]

i understand you

well goodluck with your tests, i gtg

[proper crystal]

alr seeya
thanks

[grizzled junco]

YOU VIBE CODED THE WHOLE THING?

[proper crystal]

yeah, but as I said before, I don’t really trust it

writing my own implementation as we speak

[brazen swallow]

I've just finished hand writing the simulation into cpp + with rendering for debug

[grizzled junco]

https://www.kaggle.com/code/arandomhobbyist/orbit-wars-with-the-barbell-strategy

Like any limitations here

In my code

Im thinking of making planets robust, only expending lots of ships for high production orbiting targets

Barbell strategy

Is this idea feasible or not

[grizzled junco]

What's cpp+ and what model are you using

[brazen swallow]

cpp + rendering*

I don't have a model yet, just rule based stuff

I wanted to rewrite the enviroment before starting models

[proper crystal]

I can’t speak for strategy but performance wise i’m personally trying to avoid loops when i can
the same approach might be able to speed things up for you

tho idk how much that would help as from what I understand you’re not exactly using an rl strategy

i’m currently at the environment setup phase moreso than the agent creation one
I plan on using RL with JAX so I want everything set up nicely before I start

[grizzled junco]

But i dont know how to use barbell strategy and rl

Like i have robust planets

But my attacks are always so terrible

Thats why I lose most times

And I thought rl allows for more dynamic and unconventional attacks

But im not sure

[proper crystal]

I mean
potentially
the hard part is getting things set up right

[proper crystal]

also why would you want to combine the two
genuinely curious

[proper crystal]

I'm currently looking at your notebook, and in my very unprofessional and unqualified opinon, it seems that you're a bit too conservative

[bronze pewter]

code is very short as well imo

[proper crystal]

yeah maybe more attacks would be a good idea
either I'm not running things right, or, when put up against the baseline nearest planet sniper, it just gets overwhelmed bc the sniper just outproduces

also your agent seems to really hate being put on the edge of the map

[bold harbor]

im thinking about using rl but confused about which algortihm to use

[grizzled junco]

The idea is that you need to be hyperconservative with much of your assets and also attack lucrative high risk and high reward assets expending those you dedicated

So how can I attack more while meeting this conservative stance on ship accumulation

[grizzled junco]

Dynamically computations of moves

New strategy

[proper crystal]

i think the conservative stance on ship accumulation is the real issue
because the reward comes from the number of planets you have rather than having the best planets imo

its better to have a bunch of low production planets than a few high production ones

because it seems that your total ship product/step is pretty low

and gets lower relative to the opponent as the game goes on

[proper crystal]

so whats wrong with leaving strategy completely up to the rl agent instead?

[grizzled junco]

So ur saying that i produce less ships than the enemy and attacking broadly = more ships

[proper crystal]

yeah, because more planets means more total production

[grizzled junco]

I was thinking of it

But i noticed my code is really good at defense

But I'll figure out a way to accomodate

So improvements: frequency of planets

Better orbital computations

[proper crystal]

like the counter to stockpiling ships on one planet via that planet's production is just take many planets and just outproduce the defending planet

[grizzled junco]

Hmmmmm🤔 yet what about planets with low production

Those dont yield much

[proper crystal]

but low production, low defense planets are worthwhile taking

[grizzled junco]

But I cant understand why my planet has like 700 ships im each but the opposing player sometimes gets more ships

[proper crystal]

maybe not taking low production high defense

[grizzled junco]

How so

Like i thought I had all the high production planets in the world

I'll revise the algorithm

More attacks, high production

But what about defense

Its not like you can take every planet

And how to replace the for loops

[grizzled junco]

You win by taking the most planets????

I thought that meant having the most ships

[proper crystal]

isn't the win condition whoever has the most planets after 500 steps

[grizzled junco]

Then it changes everything

[grizzled junco]

I low-key thought it was ships

Let's check the official rules

[proper crystal]

yeah im checking as well

Scoring and Termination

The game ends when:

Step limit reached: 500 turns.
Elimination: Only one player (or zero) remains with any planets or fleets.

Final score = total ships on owned planets + total ships in owned fleets. Highest score wins.

i think you're right

it is indeed ships

[grizzled junco]

Hence my current strategy of taking high production planets

But my defense is good

But offense is terrible

Ships fly into space and into the sun as if snipers drank too much vodka

Thats the problem

[proper crystal]

here I think I might just have a replication issue

[grizzled junco]

My computations functions did not account for home planet as rotational or so.e weird planet shooting ships into space

[grizzled junco]

??

[proper crystal]

oh wait

I can't paste images

uh

https://cdn.discordapp.com/attachments/566135903299305502/1504709416774275152/image.png?ex=6a07f959&is=6a06a7d9&hm=7656851c2ef7d84e6bfa91853e0c9a81d202c4adb39cae11b6f2e30551b205d4&

here try opening that?

so I'm putting yours against the nearest planet sniper bot

so from what I can see, your agent's planets are the most well defended by far

but I'm not sure you're winning
unless I'm reading the output wrong

and then sometimes the nearest planet sniper just takes your planets early game

can you run yours against the sniper and see if you get the same thing

bc that's what I have on my end

[proper crystal]

hopfully this image link works

[proper crystal]

you don't need to iterate through items in a loop if one step in your loop doesn't depend on the previous ones
you can process everything in parallel via python multiprocessing
or just turn everything into vectors and do vector math via numpy or jax

[proper crystal]

kinda like what I'm doing with the comet code rn

original loop implementation:

# Discretize the continuous ellipse into a dense array of points
dense = []
num = 5000
for i in range(num):
    t = 0.3 * math.pi + 1.4 * math.pi * i / (num - 1)
    ex = c_val + a * math.cos(t)
    ey = b * math.sin(t)
    
    x = CENTER + ex * math.cos(phi) - ey * math.sin(phi)
    y = CENTER + ex * math.sin(phi) + ey * math.cos(phi)
    dense.append((x, y))

my vectorized jax implementation:

# Discretize the continuous ellipse into dense array of points
num = 5000
t = 0.3 * jnp.pi + 1.4 * jnp.pi * jnp.arange(num) / (num - 1)
ex = c_val + a * jnp.cos(t)
ey = b * jnp.sin(t)
x = CENTER + ex * jnp.cos(phi) - ey * jnp.sin(phi)
y = CENTER + ex * jnp.sin(phi) + ey * jnp.cos(phi)
dense = jnp.stack((x,y), axis=1)

EDIT: Pasted the wrong snippet from the original code

[brazen swallow]

Matchup: neutralCostAgent vs productionWeighted:20
Matches: 1000
Start seed: 0
neutralCostAgent wins: 597
productionWeighted:20 wins: 401
Draws: 2
neutralCostAgent win rate: 59.7%
productionWeighted:20 win rate: 40.1%
Average steps: 277.351
Wall time (s): 43.7999
Average final ships: neutralCostAgent=1627.64, productionWeighted:20=2043.3
Average final planets: neutralCostAgent=15.73, productionWeighted:20=10.618

Finally getting to mess around with stuff

[proper crystal]

nice!

[cosmic hatch]

my rust env broke ;-;

gotta fix it

[proper crystal]

o7

[grizzled junco]

But i tried writefile submissoon.py and I couldnt get a spot on the leaderboard 😭

[grizzled junco]

Why are we using rust here

[cosmic hatch]

it makes the processing faster

[cosmic hatch]

here

i saw this suggestion and it worked

[cosmic hatch]

itz oki, keep trying ;-l

[fading elbow]

hi! are your agent rule-based, or have NN and RL involved?

[grizzled junco]

Rules based

[fading elbow]

yeah, it seems that rule-based models have a higher advantage over NN and RL. My origenal plan is to switch to NN after a few days, but now I'm considering not to switch

[grizzled junco]

Rules based ain't even ai

Or neural net

[grizzled junco]

Why does rules based rule

Cause it has a prepared good algorithm while models take forever or something

My algorithm has robust planets

But it is so bad at attacking

Like im cringing😭 😭 😭 😭

[fading elbow]

Do your agent have some kind of Weight for attakc&defense? I think adding a Weight to let the agent be more "aggressive" can help.

[grizzled junco]

Like i have a quota of ships and i prioritize high production planets

[fading elbow]

yeah, i think that could help. my approach is still rule-based, but I’m trying to make the attack/defense weight dynamic instead of fixed.

[fading elbow]

fr exp, if a planet is safe, it can become more aggressive and send ships to high production targets. If it is under pressure, it should preserve more ships or support nearby planets.

[grizzled junco]

Hmmmmm🤔

Thats actually kind of flexible 😄

[fading elbow]

that’s the idea 😄 I’m trying to avoid using one fixed aggression value. Safe backend planets can act more like supply planets, whle frontier planets should be more conservative unlss they have enough advantage

[bold harbor]

i have reached

1390 rank

from 2440

lesss goooo

[fading elbow]

congratulation!!

[sweet vapor]

Looking at the games it seems my agent wins it is exclusively when the other agent shoots ships into the void and not otherwise, that's around 650 ELO, so uh, anyone else particularities they otice at their ELO's?

just submitted my RL agent with 2x as much training so I'm hoping it's gotten any better 🥲

[cosmic hatch]

lesssgoooo

[proper crystal]

nice!

also do you guys have any theories as to why RL agents aren't performing as well as rule based agents

[fading elbow]

I think this is because each match is too short and with too little data. RL is better when dealing with long-term complex data. But for short term, rule-based is better

[proper crystal]

I see

makes sense

[grizzled junco]

And you have to manually train the bastard with a strategy you can program with static python

But i swear the attacks are the worst part

Defense is so easy

But how to bypass the for loops though

[cosmic hatch]

What's your strat for attacking?

[proper crystal]

honestly if you’re not doing rl, it’s probably not worth it

[grizzled junco]

High production planets

Compute distance rotational to compute distance and angle of attack

Ships needed exceeds safety floor and planet ships

[brazen swallow]

How long do the bots get per step to run?

[vivid tide]

does someone made a public rust version of the environment?

[grizzled junco]

Does kaggle even allow rust code

[cosmic hatch]

Yeeeee, it does

[proper crystal]

I’m curious about this too

I don’t think I’ve seen anything in the sim’s source code that has a timer

[proper crystal]

@brazen swallow from what I understand this is the config for the game
{'episodeSteps': 500, 'actTimeout': 1, 'runTimeout': 1200, 'agentTimeout': 2, 'shipSpeed': 6.0, 'cometSpeed': 4.0, 'seed': None}

so 1 sec / agent for the actions and 2 seconds for init?

[grizzled junco]

Do all planets have the same angular velocity

[proper crystal]

yeah i believe so

[misty oyster]

1 sec / agent with a 60 second bank (agentTimeout should be 60)

[proper crystal]

wait what does the bank do?

[austere patrol]

Maybe when your agent times out, it will spend the time in the bank

[brazen swallow]

My agent is crushing 1v1s but almost never winning the 4p

[fading vortex]

sameeee

[fading vortex]

are you using RL with self play?

[brazen swallow]

Pure rule based

Hovering around 900~ elo

[grizzled junco]

What's yr algorithm

[fading vortex]

like he'd tell you 😆

[brazen swallow]

He gets it

[grizzled junco]

How'd people compute the angle from a moving planet to static, and static to moving and even moving to moving

Precisely

[fading vortex]

I've explored a bunch of approaches. In order of complexity:

• Just simulate hundreds of launches in different angles and see what they hit
• Draw straight lines from future positions of the target to current position of the origin - if they have the right length for the fleet speed, calculate angle from origin
• Model the fleet as an expanding circle from the origin, and solve for tangency of expanding circle to target circle

[grizzled junco]

Well what about early game logic

Like im supposed to ideally find nearest high producing planets but i feel like im not expansive enough

[fading vortex]

I use RL

[compact plaza]

Hey, I wanted to ask if bigger bot system are causing more heuristics issues? Is a smaller bot better for Orbit Wars?

[grizzled junco]

But how are people using vectorized bumpy arrays

NUMPY I mean

[grizzled junco]

I thought of this

So basically I take the parametric equations of the planets orbit

Then I find the new angle after a guess of 10 turns

Then I get the distance

Divide by speed to get the actual time aka amount of turns

Is the difference between my guess and the actual close to 0

Then I can cinflate

And grab the x and y I had computed for my guess

Its like Epsilon delta proofs

[proper crystal]

hey do you know if we can run our models on a gpu?

or a tpu even maybe

or do they have to be cpu only

[fading vortex]

cpu, less than 2 (slow) cores allocated and 1 second per turn time limit 😢

unless your model is weirdly big you should still be able to do at least 1 forward per turn, but search is difficult

[proper crystal]

yikes…

no kernel launch overhead I guess…

[brazen swallow]

why doesnt obs expose num_agents to agents?

[pulsar barn]

because its not implemented

you can infer it from planets

[sage mountain]

hi, im new to agent competitions on kaggle

are the environment configurations fixed as default?

i mean this

Parameter    Default    Description
episodeSteps    500    Maximum number of turns
actTimeout    1    Seconds per turn
shipSpeed    6.0    Maximum fleet speed
sunRadius    10.0    Radius of the sun
boardSize    100.0    Board dimensions
cometSpeed    4.0    Comet speed (units/turn)

[brazen swallow]

Yes the constants are constant for the competition

[sage mountain]

ah thanks, i thought since it was introduced as "environment variables" it might be subjected to change per episode

[upbeat flint]

@misty oyster are you intending to make the fix suggested in this topic? https://www.kaggle.com/competitions/orbit-wars/discussion/702758 so that numba can be used?

[sage mountain]

hi, in the combat section it says

When one or more fleets collide with a planet (either by flying into it or being swept by a moving planet), combat is resolved:

All arriving fleets are grouped by owner. Ships from the same owner are summed.
The largest attacking force fights the second largest. The difference in ships survives.
If there is a surviving attacker:
If the attacker is the same owner as the planet, the surviving ships are added to the garrison.
If the attacker is a different owner, the surviving ships fight the garrison. If the attackers exceed the garrison, the planet changes ownership and the garrison becomes the surplus.
If two attackers tie, all attacking ships are destroyed (no survivors).

so if there are 3 arriving fleets on the same turn (possible with 4 players), what would happen to the fleet with the least amount of ships? instantly destroyed and never be considered?

[static wagon]

Hello,

Could you please advise how learned (updated) weights are expected to be stored between episodes? Are we supposed to train the model locally first, and then, once it is uploaded, it no longer continues learning?

[pulsar barn]

yes

[brazen swallow]

Pretty much

So the process is:

Combine arriving fleets by owner.
Sort owners by total arriving ships.
Compare only the largest and second-largest totals.
Ignore all lower-ranked owners.
If one owner survives, that survivor then reinforces or attacks the planet.

[rough bay]

Anyone else having trouble getting scores above 900? I feel like I was able to get into the 800s with a somewhat simple agent but no changes I have made from there has managed to get much above 900.

[brazen swallow]

Are you doing heuristics or rl?

I was unable to push higher than 900 with rule based stuff

[rough bay]

I haven't tried RL yet, trying to see how well I can do with rule based. I though I saw some people saying they were getting up to 1200 with rule based but that was earlier in the competition, I think scores have inflated a bit since then.

[brazen swallow]

The leaderboard does move up, but don't be disencouraged, your rule based agent will be invaluable when you move onto self play ppo or something else as a good opponent to collect samples from

[proper crystal]

hey, how good are you guys' RL agents at aiming if you guys are giving them free reign on choosing angles

[loud lynx]

you should give them free reign on choosing planets, there is no point in chosing angles

[proper crystal]

oh yeah ik
i just gave free reign over angles to see what would happen
it’s somehow not entirely terrible

[brazen swallow]

If you have high throughput it'll learn how to aim over time, but you should be thinking in terms of action space

Planet * angle * ships is incredibly large, especially because angle is a float here

[proper crystal]

yeah its basically infinite...
was just kinda messing around to see if it will learn to aim

as it turns out it does

kinda

[grizzled junco]

How is RL better tho

Do you still have to compute manual functions for executing moves or allocating ships or no