Full Restart Speeds Learning

摘要

Because many real-world problems can be represented and solved as constraint satisfaction problems, the development of effective, efficient constraint solvers is important. A solver's success depends greatly upon the heuristics chosen to guide the process; some heuristics perform well on one class of problems, but are less successful on another. ACE is a constraint solver that learns to customize a mixture of heuristics to solve a class of problems. The work described here accelerates that learning by setting higher performance standards. ACE now recognizes when its current learning attempt is not promising, abandons the responsible training problems, and restarts the entire learning process. This paper describes how such full restart (of the learning process rather than of an individual problem) demands careful evaluation if it is to provide effective learning and robust testing performance.