Research
We work in seven areas, and the same question runs through each of them. How sure is a learning system entitled to be about how sure it is?
Environments and graders come first, because an agent learns whatever its score rewards and the rest of the work depends on getting that right. The next three areas are the learning systems we study, from a single agent to a population of learners adapting to one another, and the last three are how we work on them.
Five working papers are in preparation, from programmes that cut across the areas, and each is listed under the area it belongs to.
Environments and graders
Where the reward is written down, and where it gets gamed.
An agent does whatever its score rewards, so the score has to be written with care. We work on scoring rules that still hold when an agent hunts for loopholes, and on simulated worlds where it can be trained and scored before anything real is at stake. Economists meet the same problem when they design an auction or a market for self-interested participants, and we treat their mechanism design and our environment design as one craft, because an incentive and a reward are gamed in the same ways.
Reinforcement learning
Learning by acting, in a world that moves while you learn.
An agent that learns from the consequences of its own actions has two problems that never quite go away. It has to decide how much to explore, and it has to decide what to keep from experience that has stopped being true. We work on both, in environments that keep changing while the agent learns.
World models
Rehearsing before acting, with a model that is wrong in places.
A model of the world lets an agent try a plan before it commits to one. We work on models that know where they are uncertain, and on planning with a model that is wrong in places, as all of them are.
Complex adaptive systems
Many learners adapting to one another, with no one in charge.
Markets, ecologies, protocols and populations of learning machines are systems whose behaviour belongs to the whole and to no one part. The newest of them are the economies run on a shared world computer, where cryptography and a stake stand in for trust and the whole state of play is open for anyone to read and to game. Selection acts on all of them much as a grader acts on an agent, and what it produces is seldom what was asked for. We study what emerges when learners adapt to one another, which of it was wanted, and how to tell the two apart early.
Probabilistic inference
Models that carry doubt about their own assumptions.
The models here are hierarchical, with priors on their priors, and we check them for calibration against what actually happened. Causal inference sits here too, because a model that acts on the world changes the data it will learn from next.
Simulation and experiment
Cheap worlds and real tests, with results that could change our minds.
We build synthetic worlds and agent-based simulations to test an idea before it meets real data, and we design experiments so that the result could prove us wrong.
Research strategy
Which questions deserve a year, and when to stop.
Choosing what to work on is itself a decision under uncertainty, with priors of its own, and knowing when to stop is part of it.
Five working papers are in preparation.
If you work on related questions, write to us.