Evolving Instance Specific Algorithm Configuration

One of the main underlying themes of the previous chapters has been to demonstrate that there is no single solver that performs best across a broad set of problem types and domains. It is therefore necessary to develop algorithm portfolios, where, when confronted with a new instance, the solver selects the approach best suited for satisfying the desired objective. This process can then be further refined to intelligently create portfolios of diverse solvers through the use of instance-oblivious parameter tuners. However, all of the approaches described thus far take a static view of the learning process. Once a portfolio has been trained, no other refinements take place regardless the amount of new data that becomes available. This chapter, therefore, focuses on an approach that demonstrates that it is at times necessary to retrain a portfolio as new instances become available. This is done by identifying when the incoming instances become sufficiently different from everything observed before, thus requiring a retraining step. The chapter shows how to identify this moment and how by doing so efficiently we can create an improved portfolio over the train-once methodology.