Show HN: A competition for small neural networks that play strategy games
tinybrains.dev15yrs back I participated in "Google Ants AI Challenge 2011", an ai programming competition, hosted by the University of Waterloo, and I ranked #127 (#1 in my country). The competition gave me a huge learning oppurtunity where developers across the world came to a forum and discussed various techniques.
Now, I've built a similar platform to bring back the fun of building a small neural network that can play the game well. Neural Network optimization seems to be much more fun.
Plz share your feedback to improve the platform and add more games.
Nice work, the new site looks great.
Can you give more background on the Ants game?
I didn't find it on the current site or the older one.
Was the game inspired by anything like agent-based simulations?
I'm not super interested in what the tech industry is calling "agentic" AI, but I am interested in collective intelligence, see David Ha's work in this area: - https://journals.sagepub.com/doi/full/10.1177/26339137221114... - https://neurips.cc/virtual/2024/105817
Would be cool if each ant itself could be an agent
Unfortunately the competition site is mostly down and couldn't find much about the old competition other than the participants blog articles. Do a search on "Google Ants AI Challenge - post mortem", and you get a lot of articles around the game.
Thanks for sharing the research. I tried implementing a per Ant decision making model, but gave up as the training time was much longer compared to the current baseline. I think I should rethink the idea.
Understood, thanks! We must need smaller models :)
I remember a game on Steam called Tiny Brains, great couch co-op.
Nice. I was 72nd. Working in AI research today and still making ai for games as a hobby (tfmbot.com is an ai i’m working on for my favourite board game terraforming mars).
Thanks for sharing. I remember #1 xathis had a score, big leap ahead of others. The difference in techniques in top 100 was almost the same.
From competing with the world to building a place for the world to compete—what a full-circle moment. Love the challenge: how much strategy can a tiny neural network learn? Excited to see what people build.
Let's play real stuff.
"Playing Hex and Counter Wargames using Reinforcement Learning and Recurrent Neural Networks"
Is "Hex" real stuff? I know that from university AI courses, from before current ML phase. I thought this was a toy game invented specifically to be nice for AI exercises - bounded, easy to follow, moderate branching factor, and designed to make ties impossible.
No no no it's not a game named Hex. It's a gaming genre where units move on a field of hexagons.
OMG!
Just yesterday I published my reworked GoNEAT library that implements HyperNEAT combined with phased search and backpropagation [1].
But it's kind of impossible to enter for me because of the hard pytorch requirements :( would love to see the project as a gym, so that you can run your own ANN design algorithm.
I get that most data science students still use python, but the evolutionary world is kinda in C++ and other native languages.
Anyways, great project nonetheless.
Where do you see a hard requirement? I have added support for ONNX model upload for now and would love to extend support for other formats. How you build the model is totally upto you. I don’t check anything other than format and inference time and model size.
Saw your repo and understood you question better. The requirement are now limiting Neural Networks only, not a direct algorithm implementation
Around 1997 there was something awfully similar to this doing the rounds, called AI Wars. IIRC it was code rather than weights, but it's a funny parallel. 30 years! God I'm old.
Interesting—how small can a neural network get and still make good strategic decisions? Curious whether these models can adapt to unfamiliar opponents.
Cool idea!
It would help to delete all the text on the page, and write it without AI.
For example, "model and manifest bytes together pick the class; every version also plays on Open"
Thanks for the feedback. I’ll take that as top priority.
This looks fun. How do you evaluate the networks — is it purely based on game performance, or are there other metrics like size and inference speed too?
Glad you like it. The evaluation is based purely on game performance. However each weight class is evaluated separately. Nano, micro, mini, small, large and open class.
When you submit a model it participates on both its weight class and the open class
This reminds me of John Conway's Game of Life. I wonder what surprises we could get by combining Game of Life with neural networks
Man, I remember doing this is 2011 as well. Everything some kind of hand-coded strategy. I enjoyed it a lot.
Pretty neat! I'm considering entering some models. How long will you be running these competitions?
The current season is a public beta and ends by end of the month. After that am considering 3 month seasons.
Signed up and submitted a test model. Now the real training begins!
I remember the Google Ants challenge being pretty interesting. What made you choose strategy games for the new competition instead of something simpler like classic board games?
I wanted to build games that are multiple-decision and control many things for each turn
This looks fun. How small are the networks you're aiming for?
the network size brackets are in TFA:
nano up to 16 KiB micro up to 128 KiB mini up to 1 MiB small up to 8 MiB large up to 64 MiBEach season has a different weight size restrictions. Currently open season is for a full production test.
There is also MIT Battlecode - same idea, but for human teams
Great idea to focus on small, efficient neural networks.
Reminds me of MechMania at UIUC - exciting!
It would be cool for sb to try jev on it :P although this is not text input at all
> Your class is measured, not chosen
> model and manifest bytes together pick the class
What?
How hard is it to write something like "your weight class is determined by the total size of the model and manifest" (if I understood it correctly).
Current version both sounds very AI-sloppy and is ambiguous.
The doc page [0] is even more painful to read.
Thanks for your feedback. I’ve fixed most text now. I did use AI to create the documentation earlier out of rush but now cleaned up everything.
Not all sloppy writing is AI. Quite a few humans also write ambiguously.
Also: https://screeps.com/
Thanks for sharing. The primary reason building this website is to learn small neural networks and tuning.