Stuart Russell -- Research areas
I am interested in building systems that can act intelligently in the real
world. To this end, I work (with various students, postdocs, and
collaborators) on a broad spectrum of topics in AI. These can
be grouped under the following headings:
- Foundations: Rationality and Intelligence
Provably intelligent systems based on the mathematical framework of
bounded optimality. Topics include quasioptimal control of
search and composition of real-time systems.
- The long-term future of
AI
Will the capabilities of AI systems exceed those of
humans? If they do, what does that imply for the future of humanity?
Can we couple the development of AI with guarantees that AI systems
will benefit humanity rather than destroying us? These are serious
questions that lead to deep research issues requiring our attention,
in much the same way as the containment of fusion reactions has become
a major topic in nuclear fusion research.
- Lethal autonomous weapons systems
Perhaps the most imminent risk from developments in AI comes from
the uncontrolled development of autonomous weapons, which may not be a good idea.
- Learning probability models
Topics include learning static and dynamic Bayesian networks and
related models and learning with prior knowledge. Applications include
speech recognition, computational biology, and human driver modelling.
- First-order, open-universe
probability models
First-order languages, such as
first-order logic, assume worlds composed of objects and relations.
Whereas closed-univserse languages such as Prolog and database systems
assume a fixed, known universe of objects, each uniquely named,
open-universe languages allow for the possibility of unknown objects
and identity uncertainty. As such, they are suitable for application
domains such as computer vision, natural language understanding, web
information extraction, computer security, and multitarget tracking,
where the set of objwcts is not given in advance. An open-universe
probability model or OUPM specifies a probability distribution
over possible worlds under the open-universe assumption. The BLOG
language provides a formal syntax, semantics, and inference capability
for OUPMs.
- Global seismic monitoring for the Comprehensive Nuclear-Test-Ban Treaty
The first major application of our work on open-universe probability models has been
in the area of nuclear arms control. We have developed a Bayesian monitoring system that detects potential nuclear events anywhere on Earth.
The results substantially improve on all existing monitoring systems. Our software runs at the
UN's International Data Centre in Vienna, generating daily bulletins.
- State estimation
State estimation (also known as filtering, tracking, belief update,
and situation assessment) is the problem of figuring out what state the world is in,
given a sequence of percepts. It is a core problem for all intelligent systems.
We have investigated both probabilistic state estimation
and nondeterministic logical state estimation; one current project
looks at the game of Kriegspiel, a version of
chess in which one cannot see any of the opponent's pieces.
- Decision making over long time scales
Like many combinatorial algorithms for sequential decision making, AlphaGo is able to make decisions with a horizon of 15 or 20 primitive steps.
A human life consists of about 20 trillion primitive steps (potential changes in muscle activation).
Even something as simple as going to a conference is about a billion steps.
Humans appear to manage by imposing a hierarchical structure on behavior, so that each primitive action is part of some
higher-level activity, and so on up to very high-level activities such
as "get a PhD" and "earn enough money to retire to Bali".
We study both hierarchical reinforcement learning and hierarchical planning and lookahead,
establishing fundamental results in each area and showing how to enable much faster learning and planning.
- Intelligent agent architectures
This topic combines all of the preceding topics in order to design
complete intelligent systems. We also examine general structural properties of
intelligent agents, including the connection between functional
decomposition of agents and additive decomposition of reward functions.
Some older projects (PNPACK, BATmobile, RoadWatch) are described here.