Evaluation criteria for genetically-tuned problem-solving experts
David B. Sturgill, Gautam Pant · 1999
Genetic algorithms have proven to be a powerful tool in solving computationally difficult problems. We present a technique for using genetic algorithms to learn a domain-specific problem solver. This technique uses a collection of search-based problem solvers, each using a different, genetically biased search procedure. Problem solvers compete in parallel to show that their search procedure, with its genetic bias, is the best; faster systems are permitted to pass their genetics on to the next generation. We compare two different techniques for learning based on two basic policies for evaluating the fitness of a problem solver. One policy assumes that problem solvers should be equally good at all problems. The alternative fitness estimate assumes that a search procedure may be very good at one kind of problem, but very bad at other kinds of problems. This paper evaluates the effectiveness of this learning technique and compares the two different learning policies in the domain of nonlinear planning.