The ability of individuals to learn optimal strategies for mitigation against infrequently-occurring natural hazards is explored. We report the results of two experiments in which participants are faced with the problem of learning the most cost-effective means of protecting against earthquake losses. The experiments utilize dynamic computer simulations in which participants are endowed with homes in virtual communities that are prone to periodic impacts by earthquakes. Participants can invest in measures that potentially mitigate losses from quakes but the effectiveness of these measures is initially uncertain. Over time participants have the opportunity to learn about true effectiveness both by direct experience with simulated earthquakes and by observing the decisions and experiences of other players. The data offer a pessimistic view of learning abilities; not only do participants persist in investing in mitigatin instruments that, in fact, have no ability to lower damage, but they also fail to fully invest in instruments that are highly effective. Among the mechanisms that appeared to impede learning was a tendency to mimic local group norms in investment levels (which are suboptimal) and to prematurely terminate attempts to learn. The paper concludes with a discussion of the implications of the work for both basic research on decision making in low-probability, high-consequence settings as well as prescriptive research in natural-hazard mitigation.