Blog
Notes on research, teaching, and things I’m thinking about.
Learning at the edge: fast, distributed, and under attack
Learning is leaving the data center — spreading across phones, sensors, and servers that are slow, unreliable, and sometimes hostile. That changes what a good algorithm has to survive.
A new home on the web
The site has a fresh look — and, more importantly, a workflow where adding a paper is a one-line change.
Which games can we actually solve?
Finding a Nash equilibrium is, in general, believed to be intractable. So the useful question is: which games are the easy ones — and can we make a hard game look easy?
Iterated random operators: a lens on learning algorithms
Many learning and RL algorithms are just a random operator applied over and over. That viewpoint buys you clean convergence guarantees.
Optimizing in real time: don't re-solve, warm-start
The world changes faster than you can re-solve a hard optimization from scratch. Perturbation theory says: reuse the solution you already have.
What is a fair price for a ride?
Ridehailing turned pricing into an algorithm that runs millions of times a day. Whose interests does that algorithm serve — and can 'fair' be made precise?
Deciding together without seeing everything
When many agents share a goal but each sees only part of the world, when does an optimal joint strategy even exist? Team theory answers that — and it's quietly the math behind multi-agent AI.
Reinforcement learning where the state space never ends
Textbook RL lives in small, tidy worlds. The real one is continuous, constrained, and only lets you learn from data you already collected. Here's some of what it takes to bridge that gap.
Designing markets for a grid that runs on weather
Today's electricity markets were built for generators you can switch on. Wind and solar answer to the weather instead — so the market itself has to be redesigned.
Security is a game: strategy against an intelligent adversary
A smart attacker reasons about your defense. That single fact turns security from an engineering problem into a game — and game theory tells you how much information to give away, and how to spread...
How do you know your self-driving car hasn't been hacked?
A hijacked sensor doesn't announce itself. The defense is statistical: make the system's own behavior betray the intruder.
Teaching cars to save fuel: reinforcement learning on the road
A connected car can see the traffic light before you can. Used well, that foresight turns into real fuel savings — and a case study in reinforcement learning that has to run in milliseconds.