Concluding Remarks

This paper has shown the significant benefits that collective memory provides to a community of agents in a problem-solving setting. We presented illustrative examples of how cooperative procedures that are beyond the scope of the scratch planner can be learned and stored into collective memory. Empirical results confirmed the usefulness of collective memory in an implemented test-bed system. These results showed that cooperative procedures and another collective memory structure, operator probabilities trees, each independently lead to reductions in the amount of time a community takes to solve randomly generated problems. Furthermore, the two structures are more effective together than either alone, showing that the structures facilitate non-overlapping aspects of learning.