Using Automated Algorithm Configuration for Parameter Control
When improving the performance of systems and solvers, we often need to consider the values for various parameters the system depends on, such as the mutation rate in an evolutionary algorithm. For each parameter, we could set a static value for all states in a system (i.e., static/offline algorithm configuration), or we could assign different values for different states (i.e., dynamic algorithm configuration). The latter option is more flexible and can potentially lead to significantly better performance. However, it is also much more challenging.
In this work, we automate this process by using an automated offline algorithm configurator. We first demonstrate that a naive application of the configurator does not produce good result at all. We then propose new techniques to significantly improve the performance of the approach. Our approach is able to consistently outperform the default parameter control policy of the benchmark derived from previous theoretical work on sufficiently large problem sizes.
Keywords
Automated Algorithm Configuration, Parameter Tuning, Optimisation
Staff
[Nguyen Dang]{nttd}