next up previous
Next: Bibliography Up: Learning from Experience: Case Previous: Indexing and similarity

Subsections

Discussion and Conclusions

CIGAR makes the assumption that similar problems have similar solutions. According to the schema theorem, genetic algorithms process syntactic string similarities and we have shown that storing and injecting syntactically similar solutions leads to increased performance with experience. CIGAR thus combines the strengths of GAs and CBR and avoids the difficult problem of measuring similarities among problems.

Our sample results indicate that CIGAR learns to increase performance at related tasks as it gains experience. We show that the simple syntactic similarity measure of hamming distance works well on combinational circuit design problems. Several general trends can be picked out. 1) The more similar the problems the larger the difference in performance between CIGAR and RIGA. 2) The more similar the problems the more individuals that can be injected without loss of performance. However, injecting too many individuals often leads to premature convergence. Since we do not usually know problem similarity injecting between $5\% - 15\%$ of the population size strikes a good balance. 3) There are several ways to choose individuals to be injected with little difference in CIGAR performance. We can inject the closest (individuals in the case-base) to the best individual in the population, the farthest from the worst, and probabilistic versions of both. In the spirit of genetic algorithm research we prefer the probabilistic versions.

We are currently applying CIGAR to problems in object identification, spectroscopic analysis of dense plasmas, and real-time targeting and re-targeting. Because these problems are computationally intensive or require real-time responses, we will be building and investigating a parallel CIGAR implementation, different selection schemes, and different case-base implementations. On the theoretical side we are modeling and quantifying our qualitatively derived parameter values for CIGAR.

Acknowledgments

This material is based upon work supported by the National Science Foundation under Grant No. 9624130.


next up previous
Next: Bibliography Up: Learning from Experience: Case Previous: Indexing and similarity
Sushil Louis
2001-10-24