pg

@Dadojvk @Trinkle23897 Personally I used validation on multiple seed after a solution achieved a new best, and our prompt explicitly prevent the agent from using any neural networks !

@Trinkle23897 How do you avoid your agent collapsing to a decision tree? Do you have a skill that explicitly disables all non-neural net learning approaches also?

Very exciting to see the cool result! Same pattern, different physics: Codex-grown heuristics matching or beating DRL agents in fluid dynamics, while staying readable, maintainable, and transferable. Heuristics were not dead. They were under-maintained.

5/5 We give much more details in our blog post donsetpg.github.io/#/blog/heurist…. We will release all the prompts, repo, and more experiments very soon!

4/5 Finally, those heuristics react much better to « transfer learning », i.e., applying the agents to more complex environments. For example, the agent successfully reused a strategy for a case at higher Reynolds, or from 2D to 3D, which is still a very active research topic in the subfield of DRL for physics.

1/5 TLDR; We used Codex to discover and maintain heuristic learning for hard fluid dynamics control cases. I’ve been applying DRL and GNN to physics since 2019,, and over the past 3 months I’ve been toying with the idea of using agents in our processes. Inspired by the blog post from @Trinkle23897, I decided to use the same strategy and have agents find readable control strategies. This means a lot to our field, where interpretability can be key for industry.

new paper at @icmlconf 🎊 We improve several gnn approaches for CFD simulations, in the same spirit as our previous iclr paper from last year! paper: arxiv.org/pdf/2605.01542 code: github.com/DonsetPG/graph…

obviously you can steer the agent during run to focus more on some parts

very small week end side project, forked from autoresearch @karpathy : same idea but using multi armed bandit scores to optimize multiple objectives at the same time repo: github.com/DonsetPG/autor…

@kfountou @PetarV_93 yeah absolutely, idea actually comes from doing differential calculus on graphs !

@pg_dons @PetarV_93 Thanks, interesting, you also found a "good" dataset for your app. Will cite.

Likely my last paper on GNNs: “Learning to Execute Graph Algorithms Exactly with Graph Neural Networks” It has been a few years since @PetarV_93 popularized the idea of neural algorithmic reasoning, which, focused on GNNs, is essentially about the ability of neural networks to learn to execute graph algorithms. In this paper, we provide a first exact learnability result for this idea (after training with gradient descent). In particular, assuming bounded degree and finite precision, we show positive exact learnability results for graph algorithms implemented in the LOCAL model, such as flooding, breadth-first search, depth-first search, and Bellman–Ford. A few details: We train on local instructions: binary vectors describing a single node’s local update and message. We train an ensemble of MLPs on these instructions and then reuse the learned MLP as the shared update rule in message passing. Essentially, we convert an algorithm into data and then overfit to this data, which guarantees correct execution for any input, up to bounded-degree and finite-precision limitations. Using NTK theory, we show a small instruction dataset can be learned exactly. Averaging independently initialized MLPs concentrates around the NTK predictor; after thresholding, the learned local rule matches the true rule with arbitrarily high probability. Guarantees (informal): any LOCAL-model algorithm running L rounds on max-degree D graphs (bounded state+message memory) can be simulated by a GNN in O(L) iterations. We also give concrete bounds for Message Flooding, BFS, DFS, and Bellman–Ford (data size/width/ensemble sizes). Limitations: bounded degree, finite-precision IDs, and an NTK/infinite-width analysis (ensembles can be large).

TIL that chrome allows you to load a local js file to override a website's one. Tested it on @KodubDev's amazing Polytrack, it works great

Romain Bardet peut savourer : il enfile le premier maillot jaune de sa carrière sur le Tour de France ! 💛 #TDF2024

Qui sont les hurluberlus dans cette histoire ? L'édito de @JocelynChavy alpinemag.fr/les-glaciers-d…

🔥 A New Era 🔥 2023 mass loss just published for all Swiss glaciers. >4% of the remaining ice was lost this year ❗️ Combined with 2022 this is 10% in 2 years ‼️ The worst two ever, in a row... 🔥 10% 🔥 🔥 2 years 🔥 @glamos_ch @matthias_huss @scnatCH scnat.ch/en/uuid/i/b8d5…

'Glacier funeral' ceremonies are increasing but yesterday's one at Sarennes Gl. was partially prevented from taking place because @alpedhuez didn't like the publicity! 👎 Sarennes was 1st 🇫🇷 summer ski glacier in 1970 and is now... gone! Pic @INRAE_France (1906-2016) 1/

🔁 RT if you think 🇫🇷 @ThibautPinot is the Super-combative of the #TDF2023! 🔁 RT si vous pensez que 🇫🇷 @ThibautPinot est le Super-combatif du #TDF2023 !

If you're an #animation studio that wants to simplify its process and save precious time on creatives projects👇 Here's an article we created in collaboration with @awscloud team that can be a game-changer for you!😺 aws.amazon.com/fr/blogs/gamet…

🧵 Some pictures of this past days “runs”: 1. Juratind