Enhancing Parameter Control Policies with State Information

Parameter control and dynamic algorithm configuration study how to dynamically choose suitable configurations of a parametrized algorithm during the optimization process. Despite being an intensively researched topic in evolutionary computation, optimal control policies are known only for very few cases, limiting the development of automated approaches to achieve them. With this work we propose four new benchmarks for which we derive optimal or close-to-optimal control policies. More precisely, we consider the optimization of the \LeadingOnes function via RLS$_{k}$, a local search algorithm allowing for a dynamic choice of the mutation strength $k$. The benchmarks differ in which information the algorithm can exploit to set its parameters and to select offspring. In existing running time results, the exploitable information is typically limited to the quality of the current-best solution. In this work, we consider how additional information about the current state of the algorithm can help to make better choices of parameters, and how these choices affect the performance. Namely, we allow the algorithm to use information about the current \OneMax value, and we find that it allows much better parameter choices, especially in marginal states. Although those states are rarely visited by the algorithm, such policies yield a notable speed-up in terms of expected runtime. This makes the proposed benchmarks a challenging, but promising testing ground for analysis of parameter control methods in rich state spaces and of their ability to find optimal policies by catching the performance improvements yielded by correct parameter choices.