期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2007
卷号：XXXVI-5/C55
出版社：Copernicus Publications
摘要：The demand for civil navigation systems in harsh environments has been growing over the last several years. The Global Positioning System (GPS) has been the backbone of most current navigation systems, but its usefulness in downtown urban environments or under heavily treed terrain is limited due to signal blockages. To help bridge these signal gaps inertial navigation systems (INS) have been suggested. An integrated INS/GPS system can provide a continuous navigation solution regardless of the environment. For civil applications the use of MEMS sensors are needed due to cost, size and regulatory restrictions of higher grade inertial units. The Kalman Filter has traditionally been used to optimally weight the GPS and INS measurements, but when using MEMS grade sensors the tuned parameters are not always the optimal ones. In these cases, the position errors during loss of the GPS signals accumulate faster than the ideally tuned case. To help correct imperfect tuning, a reinforcement learning algorithm was used to tune the Kalman filter parameters as navigation data was collected. Tuning any Kalman filter is a difficult task and is often done before navigation with the aid of the filter designer. This process often entails much iteration using the expertise of the designer, and is in no way guaranteed to result in optimal parameters. Reinforcement learning is an intelligent solution to this problem which uses a combination of dynamic programming and trial and error exploration to develop a set of optimal parameters. In comparison to a typical iterative approach, it was found that using reinforcement learning led to slightly better estimates of the tuning parameter values; furthermore, the tuning process was performed with significantly less iteration, in comparison to an exhaustive search, due to the learning capability of the method. This benefits both static parameters as well as time varying parameters since the method is capable of constantly adapting the tuning based on collected navigation data
