Stillmark/Amboss Research Updates 2024

[VSingh48]

No description provided.

[Mountlex]

Hi @VSingh48,

thank you for suggesting these papers. However, I am not sure if they fit the webpage well. In particular, as far as I have seen, these papers do not mention related work on learning-augmented algorithms or algorithms with predictions. Can you elaborate why these papers do fit to the other papers?

[VSingh48]

Thanks for the response! We have been independently working on machine learning research with regards to the Lightning Network for the last 3 years, not specifically coming from the field of algorithms with predictions as academics (most of us come from a graph ML background) however we found later that our worked converged nicely with this subfield when looking around online and finding the webpage.

In our Channel Balance Interpolation Paper, the Lightning Network presents a real life pathfinding problem on a bidirected graph in which the capacity on each side (of a given channel/bidirectional edges) is unknown but the total capacity of an edge is known, and thus trial and error must happen in case paths are selected with depleted liquidity. We created a machine learning model to predict the balance split here and realized we can leverage the model's predictions in the cost calculation of our pathfinding algorithms (Dijkstras, or even Min Cost Flow Solvers). In our domain this means that Lightning Network pathfinding algorithms can be made more reliable with model predictions by penalizing paths that contain directed edges likely to be depleted of liquidity (shown on page 7). We have productized this system and have customers in trial right now. We plan on communicating this to either the LoG conference or a workshop this year, and it has been received well from the Lightning community over the last year.

In the Bayesian Binary Search paper, probing a neighboring channel in the Lightning Network presents as a classical bisection algorithm with a significant cost in time and liquidity. We realized that converting our channel balance prediction model into a probabilistic model (akin to a distributional prediction) allowed us to bisect in probability density space and improve the performance of the search (shown on page 4). The first part is similar to fitting a surrogate function in Bayesian Optimization, however this case uses the surrogate for a binary search problem rather than optimization of a given quantity. We further realized that this could be extended not just for supervised probabilistic machine learning but also for unsupervised learning techniques that would also allow for search space density estimation. We also deployed this in production, and are communicating this to ML conferences now as well.

I believe these would be a good addition to the page as examples of research derived from real world algorithm deployments that have found to be optimized with model predictions.

[Mountlex]

Alright, I will add your contributions to the list. However, I will slightly adapt the labels to indicate that these papers are mainly use experimental evaluations. Thank you for explaining your work!