First principles planning in BDI systems

BDI (Belief, Desire, Intention) agent systems are very powerful, but they lack the ability to incorporate planning. There has been some previous work to incorporate planning within such systems. However, this has either focussed on producing low-level plan sequences, losing much of the domain knowledge inherent in BDI systems, or has been limited to HTN (Hierarchical Task Network) planning, which cannot find plans other than those specified by the programmer. In this work, we incorporate classical planning into a BDI agent, but in a way that respects and makes use of the procedural domain knowledge available, by producing abstract plans that can be executed using such knowledge. In doing so, we recognize an intrinsic tension between striving for abstract plans and, at the same time, ensuring that unnecessary actions, unrelated to the specific goal to be achieved, are avoided. We explore this tension, by first characterizing the set of “ideal” abstract plans that are non-redundant while maximally abstract, and then developing a more limited but feasible account in which an abstract plan is “specialized” into a new abstract plan that is non-redundant and preserves abstraction as much as possible. We describe an algorithm to compute such a plan specialization, as well as algorithms for the production of a valid high level plan, by deriving abstract planning operators from the BDI program.