A connectionist model of planning via back-chaining search

Great advances have marked the progress of AI planning research over the past few years. Recent systems can quickly solve problems that are orders of magnitude harder than those tackled by the best previous planners. However, we are still a long way from understanding how humans plan. Understanding how humans plan is important if we are to develop intelligent planning systems capable of dealing with complex real world problems involving uncertainty, incomplete knowledge, limited resources, non-deterministic actions and probabilistic effects. In this paper we propose a neurally plausible model of cognitive planning that exhibits a state-space search behavior. The system is based primarily on two cognitive functionalities, namely, episodic memory and perception. In spite of its simple structure, the model can search for and execute plans involving an arbitrary number of steps. The model demonstrates how general purpose cognitive faculties such as episodic memory, semantic memory and perception can be harnessed to solve a restricted subclass of planning problems. Although the schema discussed in this paper is limited in a number of ways, we discuss how it can be extended into a more powerful and expressive planning system by incorporating additional control and memory structures.