Coping with Uncertainty in Map Learning

• Authors: Kenneth Basye;Thomas Dean;Jeffrey Scott Vitter
• Venue: Machine Learning
• Year: 1997

In many applications in mobile robotics, it is important for a robotto explore its environment in order to construct a representation ofspace useful for guiding movement. We refer to such a representationas a map, and the process of constructing a mapfrom a set of measurements as map learning. Inthis paper, we develop a framework for describing map-learningproblems in which the measurements taken by the robot are subject toknown errors. We investigate approaches to learning maps under suchconditions based on Valiant‘s probably approximately correct learning model. We focus on the problem of copingwith accumulated error in combining local measurements to make globalinferences. In one approach, the effects of accumulated error areeliminated by the use of local sensing methods that never mislead butoccasionally fail to produce an answer. In another approach, theeffects of accumulated error are reduced to acceptable levels byrepeated exploration of the area to be learned. We also suggest someinsights into why certain existing techniques for map learningperform as well as they do. The learning problems explored in thispaper are quite different from most of the classification andboolean-function learning problems appearing in the literature. Themethods described, while specific to map learning, suggest directionsto take in tackling other learning problems.