EKSL's research includes work on Artificial Intelligence planning,
particularly on agents (autonomous entities) acting in an environment.
Therefore, our concerns comprise sensing, monitoring and acting as well as
classical planning.
We believe in experimental validation of our ideas, and so we have
built a number of testbeds, such as Phoenix
and TransSim.
We have also implemented tools to help define, run, and analyze
experiments. These are called CLIP and CLASP, which are described
in The
CLIP/CLASP Project.
My research is to build a substrate for planning testbeds like Phoenix
and TransSim. These testbeds are typical of many planning testbeds,
such as (sorry that this is still incomplete):
- Truckworld
- Truckworld simulates trucks with robot arms that transport goods between
cities in a graph. The trucks sense and act by grasping things with their arms.
Sensing and acting take time and can fail, requiring robust planning that can
deal with uncertainty. For more information, get on the
truckworld-users mailing list by sending mail to
truckworld-users-request@cs.washington.edu.
- MICE
- The Michigan Integrated Coordination Experiment is a testbed that simulates
agents moving on an abstract gridworld. The testbed allows the user to specify
the interactions between agents that occupy the same grid location. The
research emphasis is on distributed, cooperative problem solving. For more
information, contact ???
- Tileworld
- Tileworld simulates an agent that pushes tiles around on a grid and
receives points for filling holes with tiles. The tiles appear and disappear
randomly, thereby putting time-pressure on the agent and making the world
uncertain. For more information, contact ??
- DVMT
- The Distributed Vehicle Monitoring Testbed
- Trains
- Trains simulates trains moving cargo around in a rail network; it
is similar to Truckworld.
One common thread in these testbeds is that they are elaborate and took a great
deal of time to implement. One part of my work is to capture the common
elements of these simulators in a substrate, so that new domains can be
implemented relatively quickly.
Another common thread is that they all define a correspondence between the
real-time thinking of the agent and the corresponing amount of time that elapses
in the environment. This correspondence supports research in real-time problem
solving, since the environment continues while the agent thinks. Many of these
testbeds define that correspondence by simply measuring the cpu time of the
agent's computation and advancing the clock by a related amount. For example,
the default correspondence is Phoenix was 1 cpu second per 5 minutes of
simulation time. Other testbeds defined the simulation time taken by particular
cognitive actions of the agent and advanced the simulation clock whenever those
actions were computed.
I am interested in improving the measurement of real-time thinking in planning
testbeds.
Back to my home page
sanderso@wellesley.edu