Ephemeral Resource Constraints in Optimization

Constraints in optimization come traditionally in two types familiar to most readers: hard and soft. Hard constraints delineate absolutely between feasible and infeasible solutions, whereas soft constraints essentially specifyadditional objectives. In this chapter, we describe a third type of constraint, much less familiar and only investigated recently, which we call ephemeral resource constraints (ERCs). ERCs differ from the other constraints in three major ways. (i) The constraints are dynamic or temporary (i.e., may be active or not active), and occur only during optimization—they do not affect the feasibility of final solutions. (ii) Solutions violating the constraints cannot be evaluated on the objective function—in fact that is their main defining property. (iii) The constraints that are active are usually a function of previous solutions evaluated, bringing in a time-linkage aspect to the optimization. We explain with examples how these constraints arise in real-world optimization problems, especially when solution evaluation depends on experimental processes (i.e. in “closed-loop optimization”). Using a theoretical model based on Markov chains, the effects of these constraints on evolutionary search, e.g., drift effects on the search direction, are described. Next, a number of strategies for coping with ERCs are summarized, and evidence for their robustness is provided. In the final section, we look to the future and consider the many open questions there are in this new area.