An Alternative Approach for Reasoning about the Goal-Plan Tree Problem

BDI agents have goals to achieve and a library of plans that can be used to achieve them, typically requiring adopting further goals. A Goal-Plan Tree (GPT) structure can be used to naturally represent the goals of BDI agents with the required plans and subgoals to achieve them. These can be used to significantly improve an agent’s deliberation and ability to make reasoned decisions on plan selection and whether to commit to achieving a new goal. In previous work, a Petri net based approach for reasoning about goal-plan trees was defined. This paper outlines an alternative approach for performing the reasoning using constraint logic programming. In work by Thangarajah et al. [8, 9, 10], a GPT is defined to represent the structure of various plans and subgoals related to each goal for an individual agent. At each node in the tree, summary information is used to represent the various constraints under consideration. However, the amount of summary information could potentially grow exponentially with the size of the GPT [2], which could have a significant impact on an agent’s performance for larger problems. To this end, a different approach was introduced by Shaw and Bordini [5], mapping a GPT into a Petri net in such a way as to avoid the need for summary information for reasoning about positive and negative interactions between goals. In [6], the focus is on reasoning about resources using Petri nets, which are then combined into a coherent reasoning process encompassing the reasoning about positive and negative interactions from [5]. In this paper, we outline an alternative approach to reasoning about positive [9, 3, 5], negative [8, 1, 5] and resource [10, 4, 6] interactions using a constraint logic programming approach developed in GNU Prolog to define a set of constraints that are solved to generate a successful execution ordering of the plans to achieve the goals.