Unified criterion for state and action abstraction in autonomous agent

Autonomous abstraction of state and action is one of the key issues in the behavior acquisition problem of reactive agents. The paper proposes a general framework for the state and action abstraction, which is based on the uncertainty minimization of the behavior outcomes. This methodology not only unifies the two abstraction problems, but also provides a way to combine different abstraction criteria which have been used empirically in conventional works. An experimental study in the latter part suggests that our method increases the adaptability of the agents to the environment and improves the overall behavior performance