Learning Topological Structures of POMDP-based State Transition Models by State Splitting Method

T. Yairi, M. Togami, and K. Hori (Japan)


machine learning, POMDP, state transition model, mapbuilding, mobile robot


This paper proposes a method for learning topological structures of POMDP (Partially Observable Markov De cision Process)-based discrete state transition models from sequential data of agents' observations and actions. In this method, an initial state set is composed by clustering the observation vectors in the data or discretizing the contin uous observation space. Then, the local structure of the state transition model is gradually improved by repeatedly and selectively splitting the states so that it minimizes the "non-uniformity" or entropy of the state transitions. We examine the applicability of this method on the topological map learning problem of mobile robots.

Important Links:

Go Back