Jaddle

Jaddle: Saddle up with JAX

Jaddle: Saddle up with JAX

GitHub LinkedIn

Where Should the Air Detectives Stand?

Choosing 50 pollution monitors out of 470, and proving the choice is close to the best possible
optimisation
air quality
JAX
Jaddle

How to place 50 air-quality monitors among 470 candidate sites using real EPA data: relax the on/off choices into dimmers, solve a convex problem with Jaddle, and get a guarantee that no network does more than 1.3% better.

Oct 29, 2026
Brendan van Rooyen

Hiring a Referee for Your Search Results

Better retrieval for RAG in four layers, and where Jaddle fits
jaddle
optimisation
machine learning
RAG

How a retrieval-augmented generation (RAG) pipeline finds better pages: BM25 keyword search, learned term weights, embeddings, rank fusion, and a learned reranker trained by solving a batch of quadratic programs with Jaddle.

Oct 25, 2026
Brendan van Rooyen

Why a General Convex Solver Matters in the Age of Foundation Models

Or: in praise of bowls
optimisation
convexity
machine learning

Foundation models are brilliant at the hilly problems. This essay is about the bowl-shaped ones, where convex optimisation offers guarantees, explanations, speed and cheap experiments, and still performs surprisingly well.

Oct 25, 2026
Brendan van Rooyen

Step curve of instances certified against solve time on a log axis from 0.5 seconds to 1 hour. The curve rises steeply between 1 and 10 seconds, reaches 365 by 1 minute and 382 by the 1 hour limit, one short of the 383 instances in the set.

Jaddle on the hard MIPLIB LP benchmark: 382 of 383

Jaddle’s performance on MIPLIB LP Relaxations
jaddle
optimisation
linear programming
benchmarks

A look at Jaddle on hard LPs

Oct 23, 2026
Brendan van Rooyen

How Jaddle Works

A shopper, a referee, a horse saddle and a few clever tricks
jaddle
optimisation
duality
algorithms

An explanation of how Jaddle solves optimisation problems using duality and the Lagrangian.

Oct 22, 2026
Brendan van Rooyen

How Much Wobble Can You Handle?

Building a risk-limited investment portfolio, one ice-cream cone at a time
jaddle
optimisation
cone programming
finance

A friendly look at portfolio optimisation: shared risk, personal risk, the efficient frontier, and why the shape of a cone matters.

Oct 21, 2026
Brendan van Rooyen

The Goldilocks Battery

Not too big, not too small: using gradient descent to size a home battery
jaddle
optimisation
linear programming
energy

How to choose the right battery for a solar-powered site by solving many small daily plans at once and letting them tell you which way to grow.

Oct 21, 2026
Brendan van Rooyen

Ignoring the Fibbers

How a robust Kalman smoother keeps track of a moving object when some measurements are wildly wrong
jaddle
optimisation
estimation
robust statistics

A gentle introduction to Kalman smoothing, why a few bad measurements can wreck it, and how the Huber penalty puts outliers in their place.

Oct 21, 2026
Brendan van Rooyen

Herding a Thousand Springy Robots

Model predictive control for a whole fleet at once, and why a GPU is like a marching band
jaddle
optimisation
control
GPU

An introduction to model predictive control, and how solving a thousand small planning problems in lockstep on a GPU beats solving them one at a time.

Oct 21, 2026
Brendan van Rooyen

Teaching a Robot Banker to Play Fair

What happens when a computer learns from a biased past, and how a single rule can nudge it back
jaddle
optimisation
machine learning
fairness

A friendly tour of fairness-constrained logistic regression: why hiding the sensitive attribute isn’t enough, and how to measure the price of fairness.

Oct 21, 2026
Brendan van Rooyen

Max-Cut: The Art of Splitting a Room

How pretending a hard puzzle is easier gives you a near-perfect answer, and a promise about how good it is
jaddle
optimisation
semidefinite programming
graphs

A gentle introduction to the Max-Cut problem and the Goemans–Williamson trick: relax, solve, then slice with a random sheet.

Oct 21, 2026
Brendan van Rooyen

A Gentle Tour of Saddle Point Solvers

How two stubborn players, each learning from the other, end up somewhere useful
jaddle
optimisation
machine learning

An intuitive introduction to saddle point solvers: turning constraints into a two-player game, finding a Nash equilibrium, and why long-running averages can settle even when individual updates wobble.

Oct 21, 2026
Brendan van Rooyen

Small Problems? Use Clarabel. Big Problems? Use Jaddle.

Careful steps versus cheap steps, and how problem size decides the winner
jaddle
optimisation
Clarabel
GPU

A guide to choosing between a careful, high-accuracy solver like Clarabel and a cheap-step GPU solver like Jaddle, based on problem size and how many problems you need to solve.

Oct 21, 2026
Brendan van Rooyen
No matching items