A Non-Revisiting Genetic Algorithm with a Parameter-less Adaptive Mutation Operator

Description

Genetic algorithms, like many stochastic algorithms, suffer from the problem of revisits: a search location that has been visited before may be revisited. However, many real-world applications involve expensive and time consuming fitness evaluations. Thus revisits are highly undesirable for these problems. In this project, a novel non-revisiting genetic algorithm will be investigated. The new algorithm completely overcomes the problem of revisits by maintaining a dynamically built binary space partitioning tree structure.Early experimental results are encouraging and a large-scale study is essential to fully understand and evaluate the algorithm. This project will include (1) theoretical and experimental study of the properties of the algorithm, (2) detailed comparison and application to other evolutionary computation methods, (3) applying the method to a combinatorial optimization problem, and (4) applying the method to a real problem of interest to Hong Kong society and industry.