On benchmark properties for adaptive operator selection

Adaptive Operator Selection (AOS) algorithms are on-line adaptive techniques that adjust the probability of applying search operators to the current solution(s). Recent work has proposed new adaptation mechanisms and have tested them empirically on either artificial reward environments or on simple benchmark functions [4], [1], [2], [3]. In this short note we would like to reflect on the properties we should expect of AOS algorithms. In addition we also discuss which of these properties are being tested by simple benchmark functions like the omnipresent OneMax problem. As noted in [3], AOS algorithms basically consist of two components. First, one needs to define the credit assignment mechanism that specifies how operators are rewarded for their contribution to the search process. Secondly, one needs to define the operator selection rule that adapts the probabilities with which operators are selected during the search. The first component’s task is to continuously measure the performance of the search operators. The second component’s task is to select the appropriate operator based on their performance estimate. Here we focus on the second component. Successfull AOS algorithms should have at least the following properties: