Dynamic benchmark targeting

We study decision making in complex discrete-time dynamic environments where Bayesian optimization is intractable. A decision maker is equipped with a finite set of benchmark strategies. She aims to perform similarly to or better than each of these benchmarks. Furthermore, she cannot commit to any decision rule, hence she must satisfy this goal at all times and after every history. We find such a rule for a sufficiently patient decision maker and show that it necessitates not to rely too much on observations from distant past. In this sense we find that it can be optimal to forget.