I think most people judge reputation from high-level things e.g. num of accepted papers, and very few people actually read these papers. This means you can game the system with LLM generated papers with little consequences, and this makes things frustrating for everyone.

Refreshing to see posts like this compared to "we have 15 papers accepted at X" 🙌

🙌🎉

Always nice to see when simpler methods + good evaluations > more complicated ones. 👌

This has happened to me too many times 🤦‍♂️ Also doesn't help that Jax and PyTorch use different default initialisations for dense layers.

Well done & well deserved!! 🎉🎉 It has been awesome to see this project evolve from the early days.

The Edinburgh one will be back and running soon. We are just updating the website and other things. There is this form for people interested - forms.gle/DAbkpN9b4cUt...

Forgot to also add ⚡ quickstart link for people who like to experiment on notebooks: github.com/KaleabTesser...

Thanks for checking it out! 👍 Good point, there might be an interesting link between MoEs and hypernets. We used hypernets since they're simpler (no need to pick or combine experts), and maximally expressive (gen weights directly).

Lol yes, will had a .gitignore, missed it when copying things over.

🎯 TL;DR: HyperMARL is a versatile approach for adaptive MARL -- no changes to the RL objective, preset diversity, or seq. updates needed. See paper & code below!

Work with Arrasy Rahman, Amos Storkey & Stefano Albrecht.

📜: arxiv.org/abs/2412.04233
👩‍💻: github.com/KaleabTessera/HyperMARL

⚠️ Limitations (+opportunity): HyperMARL uses vanilla hypernets, which can inc. param. count esp. MLP hypernets. In RL/MARL this matters less (actor-critic nets are small), and params grow ~const with #agents, so scaling remains strong. Future work could explore chunked hypernets.

🔎 We also do ablations and see the importance of the decoupling and the simple initialisation scheme we follow.

📊 We validate HyperMARL across various diverse envs (18 settings; up to 20 agents) and find that it achieves competitive mean episode returns compared to NoPS, FuPS, and modern diversity-focused methods -- without using diversity losses, preset diversity levels or seq. updates.

💡To address the coupling problem, we propose 𝐇𝐲𝐩𝐞𝐫𝐌𝐀𝐑𝐋: a method that explicitly 𝐝𝐞𝐜𝐨𝐮𝐩𝐥𝐞𝐬 obs- and agent-conditioned gradients with hypernetworks. This means obs grad noise is avg. per agent (Zᵢ) before applying agent-cond. grads (Jᵢ) -- unlike FuPS, which entangles both.

🔬 We isolate FuPS’s failure in matrix games: shared policies struggle when agents need to act differently. Inter-agent gradient interference is at play -- especially when obs and agent IDs are 𝐜𝐨𝐮𝐩𝐥𝐞𝐝. Surprisingly, using only IDs (no obs) performed better and reduced interference.