BDI agent architecture for a POMDP planner

Abstract

Traditionally, agent architectures based on the Belief-Desire-Intention (BDI) model make use of precompiled plans, or if they do generate plans, the plans do not involve stochastic actions nor probabilistic observations. Plans that do involve these kinds of actions and observations are generated by partially observable Markov decision process (POMDP) planners. In particular for POMDP planning, researchers make use of a POMDP planner which is implemented in the robot programming and plan language Golog. Golog is very suitable for integrating beliefs, as it is based on the situation calculus and researchers have drawn upon previous research on this. However, a POMDP planner on its own cannot cope well with dynamically changing environments and complicated goals. This is exactly strength of the BDI model; the model is for reasoning over goals dynamically. Therefore, in this paper, researchers propose an architecture that will lay the groundwork for architectures that combine the advantages of a POMDP planner written in the situation calculus, and the BDI model of agency. The researchers show preliminary results which can be seen as a proof of concept for integrating a POMDP into a BDI architecture