Integration of Learned Knowledge by Structured Boltzmann Machines.
Taku Okuno, Yukinori Kakazu · TRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGINEERS Series C · 1994
This paper proposes a framework for integrating independently learned knowledge by Boltzmann machines. Auto-associative neural networks based on distributed representation have some advantageous properties of dealing with knowledge. However, as to the supplement of knowledge under a dynamically changing environment, or the merging of independently acquired knowledge, they do not work well enough because of difficulty in supplementary learning. Therefore we propose a model for integrating knowledge by connecting independently learned network modules without additional learning. First, localized representation, which is indispensable for realizing interaction between knowledge in different modules, is introduced. Although it will spoil the flexibility of distributed representation, competitive learning adopted for translating representations preserves it to some extent. Secondly, the configuration of the connected network is explicated, and is behavior is illustrated. It is completely realized by activation, competition, and inhibition of Boltzmann machine. Finally, to demonstrate the effectiveness of the proposed model, results of computer simulation on a navigation problem of obstacle avoidance are shown.