Victor-Alexandru Darvariu

Dan is truly an amazing person and I hope he will do well in office. The problems ahead are very thorny, and the threat of the far-right will linger on, but it's worth taking a moment to celebrate his victory.

Not even the Romanian diaspora in Western Europe, who counterintuitively voted overwhelmingly in favour of the Eurosceptic candidate (?!), could turn the tide.

The country rallied around Dan in a campaign that involved many Romanians sitting down with their relatives and friends, explaining the threats of far-right politics.

It was a truly strange two weeks in between voting rounds, in which Dan's opponent could not have sabotaged his leading position more if he tried (ghosting debates, ad-hominem attacks, ...).

He managed an incredible victory against his Eurosceptic, ultranationalist adversary, who earned 41% of the vote in the first round against Dan's 21%.

Dan went on to study at École Normale Supérieure and then did a PhD at Paris 13, returning to Romania afterwards as a mathematician, and eventually got into politics.

I'd heard he did olympiads in his youth but I was blown away by his accomplishments! Other 1988 gold medallists whose names you might recognise are Ngô Bào Châu and Terence Tao, both of whom went on to earn Fields medals.

In an increasingly rare W for democracy: Romania's president-elect, Nicușor Dan, is an IMO gold medallist. He participated in 1987 and 1988, and got perfect scores both times! www.imo-official.org/participant_...

Tree search in DAG space with model-based reinforcement learning for causal discovery | Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences Identifying causal structure is central to many fields ranging from strategic decision making to biology and economics. In this work, we propose Causal Discovery Upper Confidence Bound for Trees (CD-U...

You can read the full paper here: royalsocietypublishing.org/doi/full/10..... We also open source our code and data at github.com/VictorDarvar.... 8/

The method is broadly applicable to any DAG construction task. If you work on causal inference, reinforcement learning, or combinatorial optimization, we believe CD-UCT offers a promising new direction. 7/

We conduct a comprehensive empirical evaluation on both synthetic and real-world datasets. Across the board, CD-UCT consistently outperforms the state-of-the-art model-free RL approach and greedy search baselines. 6/

Our method applies broadly to causal Bayesian networks, handling both discrete and continuous random variables, which makes it suitable for a wide range of domains. 5/

A key contribution is an efficient, formally proven algorithm for excluding edges that would introduce cycles, enabling deeper and more effective discrete search during DAG construction. 4/

CD-UCT incrementally builds directed acyclic graphs (DAGs) through a targeted tree search, improving substantially over more standard model-free approaches such as RL-BIC. 3/

Identifying causal structure is fundamental to many fields including strategic decision-making, biology, and economics. In this paper, we introduce CD-UCT, a model-based reinforcement learning method for causal discovery. 2/

Screenshot of the paper title, authors, abstract and miscellaneous bibliographical information as it appears in the published journal PDF.

Our paper "Tree search in DAG space with model-based reinforcement learning for causal discovery" has just been published in Proceedings of the Royal Society A. Joint work with Steve Hailes and @mircomusolesi.bsky.social 🧵 1/

Here's to a fresh start!

I've used www.sky-follower-bridge.dev and github.com/marcomaroni-..., both tools are pretty stable!

Just done migrating over to greener pastures.

Feels great to be recognised together with my colleagues after a pretty intense reviewing season! 🫡 https://x.com/LogConference/status/1862602407395697123

Thread with an overview of the paper:

Super happy to see this work published in TMLR (with the survey certificate). We had a great discussion and experience with the venue overall, can't recommend it enough in comparison with the usual ML conference lottery. https://x.com/TmlrPub/status/1829141995115155541

I will greatly miss collaborators, colleagues, and friends that have shaped my thinking throughout this time. I wanted to take the opportunity to say a big thank you to all. 4/4

Career update: I recently joined @oxfordrobots as a postdoc to work wih @hawesie and his team. 🧵 1/4

This work has emerged after a few years of dealing with these problems in practice, and I hope it offers a fresh perspective that people might find useful. Feedback is most welcome! 10/10

That's the gist of it! Given the scope, it's quite possible that I've missed some references, so feel free to contact me with works you think should be included. 9/10

Additionally, we give a concise introduction to RL algorithms and graph representation learning techniques. These areas evolved largely independently but are commonly paired in Graph RL. 7/10

The methods used in these works share substantial commonalities despite having been proposed in different fields. We therefore bring them together under the "Graph Reinforcement Learning" framework. 5/10

Various works in the literature have been turning to reinforcement learning (RL) as a flexible way to solve optimization problems by trial-and-error. 3/10

New pre-print now on arXiv, Graph Reinforcement Learning for Combinatorial Optimization: A Survey and Unifying Perspective (https://arxiv.org/abs/2404.06492). Joint work with Steve Hailes and @mircomusolesi. 🧵 1/10