We present experimental results for a new algorithm for path planning under
uncertainty (stochastic shortest path problems). Details of the algorithm
described here can be found in the paper submitted to UAI2001.
Usually in MDPs, the objective function is to minimize the expected cost.
In robotics, however, one is often interested in finding feasible paths to
the goal. In problems with uncertainty, this translates to finding paths with
high probability of success. Of the paths with high success probability, we
would like to pick the one that is expected to be shortest. In the paper, we
illustrate the effects of this tradeoff further.
Here, we apply the algorithm to several test cases. These test cases involve uncertainty in the position of obstacles. This type of uncertainty could come from noisy sensors data, for example. The noise in these examples is gaussian and the shaded areas are a graphical representation of this distribution.
Combination of Local Controllers;
Carlos Guestrin and Dirk Ormoneit;
To appear in 17th
Conference on Uncertainty in Artificial Intelligence (UAI-01), Seattle,
Washington, August 2001.
version of presentation]