Can Search Play A Role in Practical Applications ?

Received wisdom has it that search is ineeective in elded AI systems; it is argued that realistic problem spaces are simply too large to allow eeective search. We suggest that this is because search-based techniques fall prey to early mistakes, making fatal errors suuciently early in the search that backtrack-based techniques are incapable of correcting them. We describe four general techniques that have been used in practical applications to circumvent this diiculty: limited discrepancy search, schedule packing, squeaky wheel optimization, and relevance-bounded learning. We explain how each technique avoids the early mistakes problem, and describe the impact of the technique on speciic problems of practical interest.