Learning of information gathering in modular intelligent systems

摘要: Research of intelligent systems aims to realize autonomous agents capable performing various functions ease every day life humans. Usually, such occupations can be formalized as a collection tasks that have executed in parallel or sequence. Since real world environments are highly dynamic and unpredictable, require cognitive capabilities learn how execute through interactions. Considering the system has limited resources for acquiring processing information, strategy is required find update task-relevant information sources efficiently time. This thesis proposes level approach gathering process an implementation puts this idea into work. The presented framework takes modular where modules defined elementary units acquisition processing. design helps handling scenario complexity. A module management mechanism learns which deliver task relevant constrained distributed among these reward based framework. This reduces partial observability caused by provides better support other high functionalities system. Such adaptive also makes it possible deal with variations environment. Two different applications simulation implemented test hypotheses demonstrate utility proposed framework: first implements `reaching-while-interacting' humanoid robot second employing navigation mobile robot. Both scenarios involve objects, rendering challenging environment close real-world conditions Results from experiments provide evidence postulated thesis.