Acquiring plans within situated, resource-bounded agents: a hybrid, BDI-based approach

© 2009 Dr. Samin M. R. Karim.The BDI model is a widely accepted model for situated, resource-bounded agents. Intentions are a key component of the BDI model, and constrain the agent’s commitment to achieving its desires, and execution of behaviours. These behaviours are fulfilled via plans, which are an abstract representation for behaviours. Plans specify a course of action in a given context to achieve a particular goal. Acquiring plans is a complex problem, which is due to the complex relationships between the plan goal/context and the presence of multiple action steps in a plan. The BDI model does not feature a plan acquisition, or more generally, a knowledge acquisition, capability. Acquiring atomic knowledge is comparatively more straightforward, as reinforcement and management of the knowledge is less complex than plans. This thesis presents an approach to plan acquisition for agents based on the BDI model and combined with a ‘bottom-level’ learner. The system, which we call PGS (plan generation system), essentially transforms knowledge from the bottom-level into the BDI-based top-level. PGS manages the top-level and bottom-level interactions whilst managing action execution. These processes all occur during run-time. We will firstly explain concepts that are central to the thesis, and then describe related work that goes towards achieving the thesis aims. We then describe the PGS architecture, the results from case study experiments that have been conducted, and a discussion of these results and the main thesis outcomes