Incremental Rule Chunking for Problem Solving

In this paper we address the issues of how incrementally chunking learned action rules of increasing length and complexity can assist in solving problems of ever greater complexity. To this end, we employ a micro-world with simple objects and simplified physical behaviors. The agent first learns some basic elemental rules capturing the fundamental physical behaviors of the agent itself, the objects and their interactions. Then, some moderately complex problems such as going from a start state to a goal state that do not require too many steps are given to the system and the system uses a standard search process (e.g., A) to find solutions which do not require too much search time because the problems are relatively simple. The solutions are then remembered as "chunked" rules of taking a sequence of actions to achieve a certain goal. Later, when a more complex problem - one that requires many steps to solve - is encountered, the chunked rules discovered earlier can be used to greatly reduce the search space by providing chunked sub-steps. Problem solving for complex problems without the chunking process would be impossible, as the search space would be combinatorially large.