State-Space Modeling Using Self-Organizing Maps

摘要

Dynamic processes can be modeled using self-organising maps, for example, by using the states of the system as input feature vectors to the self-organisation algorithm. The neural net converges according to the distribution of the state variables, and, in principle, this map can be seen as a compressed version of the state space. However, the distance metric which the Kohonen type adaptation algorithms are based on is not justified when measuring distances between process states. In this paper, the process to be modeled is modified so that the 'static' norm between states is justified also when measuring distances between trajectories.