基于加强学习的BDI Agent动作规划机制

摘要

To enable BDI agents making rational decisions about an appreciate course of action to pursue goals in dynamic and complex environments, the intention decision structure is described with the and/or graph which integrates goals with their opposite plans. Then,three different types of action planning decision strategies are developed based on the reinforcement learning according to the intention decision structure, respectively, single-step planning of the short-sighted BDI agents, multi-steps planning of the far-sighted BDl agents and the optimal-planning of the ideal BDI agents. Compared with traditional BDI agents systems, this new intention decision scheme overcomes the limitation of abstract high-level plans and easily to be implemented.%为了确保BDI Agent在动态,复杂的环境中实现基于某目标的动作序列决策任务,使用与/或图描述了意图决策结构,此结构将目标与实现这些目标的计划联系起来.根据意图决策结构,提出了3种不同的基于加强学习的动作规划策略,分别是短视性BDI Agent的单步规划、具有远见的BDI Agent的多步规划和追求完美的BDI Agent的最优规划.与传统的BDI Agent系统相比,这种新的意图决策模式克服了计划抽象的不足,并且易于编程实现.