Optimal experimentation in signal-dependent decision problems

The literature on experimentation and learning typically imposes a special dynamic structure: The only connection between periods is the updating of beliefs. Hence, both the present action and present signal realization only affect the future by changing the distribution of future beliefs. In many dynamic problems, however, either the present action or the present signal realization may enter future payoffs directly and not just through future beliefs: This is called signal dependence. We analyze optimal experimentation, under signal dependence, and show that experimentation may reduce information. We also provide sufficient conditions on the primitives for information-increasing experimentation.