Victor-Alexandru Darvariu
@vadarvariu.bsky.social
postdoc @ oxford robotics institute. interested in reinforcement learning, graphs, robots, and combinatorial optimization.
https://victor.darvariu.me
https://victor.darvariu.me
While exploring this lovely city I bumped into a mobile sculpture by Alexander Calder. I'm not sure whether it's the one on the cover of the classic CLRS Introduction to Algorithms textbook, but it must be a closely related cousin at least!
August 18, 2025 at 6:46 PM
While exploring this lovely city I bumped into a mobile sculpture by Alexander Calder. I'm not sure whether it's the one on the cover of the classic CLRS Introduction to Algorithms textbook, but it must be a closely related cousin at least!
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/
April 28, 2025 at 11:13 AM
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/
Career update: I recently joined @oxfordrobots as a postdoc to work wih @hawesie and his team. 🧵 1/4
April 25, 2025 at 5:28 PM
Career update: I recently joined @oxfordrobots as a postdoc to work wih @hawesie and his team. 🧵 1/4
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
April 25, 2025 at 5:28 PM
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
We also had some great discussions about the work and ML in OR more broadly, on what turned out to be an unusually sunny day in Edinburgh! 11/11
April 25, 2025 at 5:27 PM
We also had some great discussions about the work and ML in OR more broadly, on what turned out to be an unusually sunny day in Edinburgh! 11/11
Furthermore, we propose the use of decision-time planning methods (i.e., Monte Carlo Tree Search) as a way of alleviating the large training costs and poor scalability of existing model-free RL approaches. 6/11
April 25, 2025 at 5:27 PM
Furthermore, we propose the use of decision-time planning methods (i.e., Monte Carlo Tree Search) as a way of alleviating the large training costs and poor scalability of existing model-free RL approaches. 6/11
(from browsing census data, this tree shape seems representative of most London boroughs, but not for the UK as a whole: https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/articles/ukpopulationpyramidinteractive/2020-01-08)
April 25, 2025 at 5:27 PM
(from browsing census data, this tree shape seems representative of most London boroughs, but not for the UK as a whole: https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/articles/ukpopulationpyramidinteractive/2020-01-08)
The borough I just moved into is so true to its name that even its population pyramid is a tree.
April 25, 2025 at 5:27 PM
The borough I just moved into is so true to its name that even its population pyramid is a tree.
For a quick summary of the work, have a look at our poster 📜.
April 25, 2025 at 5:27 PM
For a quick summary of the work, have a look at our poster 📜.
The method we propose shows superior performance compared to a variety of existing algorithms, as well as desirable properties w.r.t. model evaluation speed and out-of-sample generalization in some cases. 5/9
April 25, 2025 at 5:26 PM
The method we propose shows superior performance compared to a variety of existing algorithms, as well as desirable properties w.r.t. model evaluation speed and out-of-sample generalization in some cases. 5/9
In this work, we formulate graph construction as a Markov Decision Process, and use Reinforcement Learning and Graph Neural Networks machinery to learn how to build networks in a goal-driven way. 3/9
April 25, 2025 at 5:26 PM
In this work, we formulate graph construction as a Markov Decision Process, and use Reinforcement Learning and Graph Neural Networks machinery to learn how to build networks in a goal-driven way. 3/9
After quite a ride, our work "Goal-directed graph construction using reinforcement learning" has been published today in Proceedings of the Royal Society A! Joint work with Steve Hailes and @mircomusolesi. 🧵 1/9 https://royalsocietypublishing.org/doi/10.1098/rspa.2021.0168
April 25, 2025 at 5:26 PM
After quite a ride, our work "Goal-directed graph construction using reinforcement learning" has been published today in Proceedings of the Royal Society A! Joint work with Steve Hailes and @mircomusolesi. 🧵 1/9 https://royalsocietypublishing.org/doi/10.1098/rspa.2021.0168
On small game instances, we match exhaustive search almost exactly, and do better than various baselines on larger games. Our imitation learned policy preserves most of the performance while being orders of magnitude cheaper to evaluate.
April 25, 2025 at 5:26 PM
On small game instances, we match exhaustive search almost exactly, and do better than various baselines on larger games. Our imitation learned policy preserves most of the performance while being orders of magnitude cheaper to evaluate.
We use Monte Carlo Tree Search to search for optimal equilibria, then use imitation learning to arrive at a Graph Neural Network policy that generalizes across game instances. We propose a very simple but efficient trick for learning this policy quickly.
April 25, 2025 at 5:26 PM
We use Monte Carlo Tree Search to search for optimal equilibria, then use imitation learning to arrive at a Graph Neural Network policy that generalizes across game instances. We propose a very simple but efficient trick for learning this policy quickly.
Formal statistical tests indicate significant differences in emotional state values when an object is present versus when it is absent, even though the sample size is relatively small. 7/8
April 25, 2025 at 5:25 PM
Formal statistical tests indicate significant differences in emotional state values when an object is present versus when it is absent, even though the sample size is relatively small. 7/8
Empirically, we find the distribution of objects present in these everyday scenes is skewed and heavy-tailed, with objects associated to indoors environments being most frequent. 5/8
April 25, 2025 at 5:25 PM
Empirically, we find the distribution of objects present in these everyday scenes is skewed and heavy-tailed, with objects associated to indoors environments being most frequent. 5/8
Happy to announce that our new pre-print, "Improving the Robustness of Graphs through Reinforcement Learning and Graph Neural Networks", is out! https://arxiv.org/abs/2001.11279 Joint work with Steve Hailes and @mircomusolesi. 1/8
April 25, 2025 at 5:24 PM
Happy to announce that our new pre-print, "Improving the Robustness of Graphs through Reinforcement Learning and Graph Neural Networks", is out! https://arxiv.org/abs/2001.11279 Joint work with Steve Hailes and @mircomusolesi. 1/8
I printed out a picture of the Father to watch over the networks training.
April 25, 2025 at 5:24 PM
I printed out a picture of the Father to watch over the networks training.
We are recruiting participants for a mobile study about images from our everyday lives and our mood. Please consider downloading our app and participating in the study! https://iss-mymood.geog.ucl.ac.uk/website/
April 25, 2025 at 5:24 PM
We are recruiting participants for a mobile study about images from our everyday lives and our mood. Please consider downloading our app and participating in the study! https://iss-mymood.geog.ucl.ac.uk/website/