Action-Level Intention Selection for BDI Agents.

Belief-Desire-Intention agents typically pursue multiple goals in parallel. However the interleaving of steps in different intentions may result in conflicts, e.g., where the execution of a step in one plan makes the execution of a step in another concurrently executing plan impossible. Previous approaches to avoiding conflicts between concurrently executing intentions treat plans as atomic units, and attempt to interleave plans in different intentions so as to minimise conflicts. However some conflicts cannot be resolved by appropriate ordering of plans and can only be resolved by appropriate interleaving of steps within plans. In this paper, we present SA, an approach to intention selection based on Single-Player Monte Carlo Tree Search that selects which intention to progress at the current cycle at the level of individual plan steps. We evaluate the performance of our approach in a range of scenarios of increasing difficulty in both static and dynamic environments. The results suggest SA out-performs existing approaches to intention selection both in terms of goals achieved and the variance in goal achievement time.