SR-APL: A Model for a Programming Language for Rational BDI Agents with Prioritized Goals

摘要

Recently, there has been much work on incorporating declarative goals in Belief-Desire-Intention Agent Programming Languages (e.g. [3]). In a BDI APL with declarative goals (APLwDG), declarative goals are used essentially for monitoring goal achievement and performing recovery when a plan has failed, performing rational deliberation, and reacting in a rational way to changes in goals that result from communication. While APLwDGs have evolved over the past few years, to keep them tractable and practical, they sacrifice some principles of rationality. In particular, while selecting plans to achieve a declarative goal, they ignore other concurrent intentions of the agent. As a consequence, the selected plans may be inconsistent with other intentions. Also, these APLwDGs typically rely on syntactic formalizations of declarative goals, whose properties are often not well understood. An Example Consider a blocks world domain, where there are four blocks, one of each color, blue, yellow, red, and green. There is only a stacking action stack(b, b'): b can be stacked on b' in state s if b ≠ b', both b and b' are clear, and b is on the table in s. Assume that the agent initially has the following two goals: Φ_1, i.e. to eventually have a 2 blocks tower with a green block on top and a non-yellow block underneath, and Φ_2, i.e. to have a 2 blocks tower with a blue block on top and a non-red block underneath. Also, her plan library has only two rules: if she has the goal that Φ_1 and knows about a green block b and a distinct non-yellow block b' that are clear and are on the table, then she should adopt the plan of stacking b on b', and similarly for the goal that Φ_2. Thus according to this library, one way of building a green non-yellow (and a blue non-red) tower is to construct a green-blue (a blue-green, resp.) tower. While these two plans are individually reasonable, they are inconsistent with each other, since the agent has only one block of each color. Thus a rational agent should not adopt these two plans. However, it can be shown that a typical APLwDG agent (that does not consider the overall consistency of her intentions) may adopt these two plans together, and may make the other goal impossible by executing one of them. The problem arises in part because actions are not reversible in this domain, a common occurrence.