A Stochastic Belief Management Architecture for Agent Control
Gavin Rens, Thomas Andreas Meyer, Deshendran Moodley · 2017
We propose an architecture for agent control, where the agent stores its beliefs and environ- ment models as logical sentences. Given succes- sive observations, the agent’s current state (of be- liefs) is maintained by a combination of proba- bility, POMDP and belief change theory. Two ex- isting logics are employed for knowledge repre- sentation and reasoning: the stochastic decision logic of Rens et al. (2015) and p-logic of Zhuang et al. (2017) (a restricted version of a logic de- signed by Fagin et al. (1990)). The proposed ar- chitecture assumes two streams of observations: active, which correspond to agent intentions and passive, which is received without the agent’s di- rect involvement. Stochastic uncertainty, and ig- norance due to lack of information are both dealt with in the architecture. Planning, and learning of environment models are assumed present but are not covered in this proposal.