Learning of the way of abstraction in real robots

Abstract

Real robots should be able to adapt flexibly to various environments. The main problem is how to abstract useful information from a huge amount of information in the environment. This is called the frame problem. The paper proposes a new architecture which can learn how to perform abstraction while executing the task. We call the architecture the situation transition network system (STNS). By this architecture, a robot can acquire a necessary and sufficient symbol system for the current task and environment. Furthermore, this symbol system is flexible enough to adapt to changes of the environment. STNS performs cognitive learning and behavior learning parallelly while executing the task. In cognitive learning, it extracts situations and maintains them dynamically in the continuous state space on the basis of rewards from the environment. A situation can be regarded as an empirically obtained symbol. In behavior learning, it constructs a MDP (Markov decision problem) model of the environment on the abstracted situation representation. This model is used for planning of behavior. The validity of STNS is shown in computer simulations.