Conclusion

Abstract

In this paper we have taken several steps towards the establishment of chunking as a general learning mechanism. We have demonstrated that it is possible to extend chunking to complex tasks that require extensive problem solving. In experiments with the Eight Puzzle, Tic-Tac-Toe, and a part of the R1 computer-configuration task, it was demonstrated that chunking leads to performance improvements with practice. We have also contributed to showing how chunking can be used to improve many aspects of behavior. Though this is only partial, as not all of the different types of problem solving arose in the tasks we demonstrated, we did see that chunking can be used for subgoals that involve selection of operators and application of operators. Chunking has this generality because of the ubiquity of goals in Soar. Since all aspects of behavior are open to problem solving in subgoals, all aspects are open to learning. Not only is Soar able to learn about the task (chunking the main goal), it is able to learn about how to solve the task (chunking the subgoals). Because all aspects of behavior are open to problem solving, and hence to learning, Soar avoids the wandering bottle-neck problem.