BDI Agent Reasoning with Guidance from HTN Recipes

Belief-Desire-Intention (BDI) agent systems and Hierarchical Task Network (HTN) planning systems are two popular approaches to acting and planning, both of which are based on hierarchical and context-based expansion of subgoals. Over the past three decades, various authors have recognised the similarities between the two approaches, and developed methods for making the domain knowledge embedded in one system accessible to the other, and for augmenting BDI agents with the ability to perform HTN-style "lookahead" planning. This paper makes a novel contribution to this strand of work by developing a formal account of "plugging" in available HTN hierarchies (e.g. from the International Planning Competition) into a BDI agent's goal-plan hierarchy. When combined with lookahead-based execution, the agent is then guaranteed to behave in accordance with the "operational guidelines" embedded in the HTN hierarchies. We also explore how HTNs could be used to obtain BDI hierarchies that can be executed without performing any lookahead. In particular, we first characterise a useful class of BDI agent hierarchies that any such translation should produce, and we then characterise the restrictions that need to be imposed on HTNs in order to encode them as useful BDI hierarchies.