Sensor-fusion with Incremental Learning Ability

Jianbo Su

Professor and Chairman of Department of Automation, Shanghai Jiaotong University

Visiting Scholar in the Center of Automation and Autonomous Systems,
Department of Electrical Engineering and Information Technology, TU Munich

Mittwoch, 20.02.2002, 10 Uhr c.t., M6-114
Frequently in practice, a multi-sensor fusion system needs to be upgraded by integrating additional sensors into the system to adapt to more complex environments and applications. Normally the structure and fusion algorithm of the fusion system should be designed from the very beginning for the upgrade, even if most of the sensors of the system are retained without any changes. This inefficiency can be overcome if the fusion system has incremental learning ability. An effective learning algorithm with incremental learning ability, called Receptive Field Weighted Regression (RFWR) algorithm has been proposed by Schaal and Atkeson. This algorithm can overcome some difficulties occurring normally in the incremental learning tasks, especially the bias-variance dilemma and the negative interference problems, thus is efficient to deal with a sufficiently complex learning task and can be expected to have wide applications in many disciplines. However, direct application of RFWR in the multi-sensor data fusion is not practical due to its computation complexity. Moreover, the idea of learning employed in RFWR is not proper for multi-sensor data fusion applications. Thus we address in this seminar how the learning algorithm in RFWR is modified to meet the characteristics of the multi-sensor fusion applications. A new cost function based on the idea of back propagation (BP) for learning in RFWR is proposed so that balanced updates among all receptive fields are reached with additional reduced computation complexity. With the new cost function, back propagation algorithm is consequently involved for learning. At the same time, all remarkable features of the RFWR, such as incremental learning ability and efficiency to approximating complex functions, are retained and improved. The modified RFWR algorithm is then served as an efficient tool for learning in multi-sensor data fusion problem. We show that the modified RFWR is inherently fit for sensor fusion problems not only in its learning ability but also in its computation structure. Combined with the weighted average strategy, a new computation paradigm is formed for multi-sensor data fusion system.

sfb-logo Zur Startseite Erstellt von: Anke Weinberger (2002-02-14).
Wartung durch: Anke Weinberger (2002-02-14).