Consumer leakage hurts both the producer and the consumers,
and therefore all players have an incentive
to counteract it. Both consumers and producers can do
this by putting more emphasis on exploration as
opposed to (pure) exploitation. Consumers could
use a variety of schemes; for example, they could choose
to subscribe at random with a non-zero probability even
if their expected surplus is less than the subscription fee,
and this probability could diminish monotonically as
the difference between these quantities increases.
Producers could fight leakage by
temporarily decreasing prices to resurrect
consumers who have mistakenly disenfranchised themselves,
in hopes that, with additional samples, the consumers will
increase their estimated surplus to levels that can support
higher prices. One might expect that consumers could be
enticed back into the market
even with small discounts, since, once their 
ventures into a realm where subscription appears to be
unprofitable, it is frozen at this just-barely-unprofitable
value 
.
Here we focus on the producer's strategy for enticing overly pessimistic consumers back into the market. (This is not to deny that consumers' exploration strategies merit serious study.) From our study of consumer leakage, it is apparent that the producer's strategy must involve dynamic pricing, and that it must cope with a profit landscape that changes dynamically due to shifts caused by consumers' ongoing attempts to learn an estimate of g. It also seems most likely that the pricing strategy would involve stochastic search on this dynamically changing landscape, rather than following some pre-planned schedule.
The dynamically changing nature of the profit landscape in our problem limits the applicability of standard stochastic optimization techniques. For example, approaches such as simulated annealing implicitly assume that the search is being conducted on a static landscape as the value of its temperature parameter is lowered. On the other hand, standard gradient-based optimization approaches are of limited use because it is often too hard to determine the gradient of the profit landscape for general g and h distributions or for different price structures. More sophisticated approaches that are currently oriented towards static landscapes might be modified to handle dynamic landscapes. In this paper, we apply a simple direct search method called the amoeba algorithm for profit maximization. The amoeba algorithm is a good candidate optimizer since it makes very few assumptions about the underlying problem domain. Although the amoeba can get stuck at local optima, our preliminary work indicates that it works well for a variety of g and h distributions, even in problems where the price structure involves as many as ten parameters [3].