Models of Cooperation, Learning and Catastrophic Risk

Abstract: Our world presents us with dangers and opportunities. Some of these dangers and opportunities are easier to handle if two or more individuals learn to cooperate. This thesis contributes five papers about cooperation, learning and catastrophic risk. In papers I-II, we consider the Finitely Repeated Prisoners' Dilemma, a model for where cooperation between two players is particularly hard to achieve. We introduce and model strategies that attempt to convince others to cooperate when backward induction can be used to eliminate cooperation for a number of steps from the end. We find that in a population with these strategies, cooperation can become recurrent, and we examine the conditions for this. Recurrent cooperation is possible in an evolutionary model (paper I) as well as in a population of players that are near-perfect Bayesian expected utility-maximizers (paper II). In paper III, we consider a bargaining model of climate negotiations where players negotiate emissions and sudden catastrophic damage occurs if emissions exceed a threshold amount. We introduce and model a mechanism of strategic reasoning, where players predict the emission bids of others, and consider how this affects the possibility of reaching agreements preventing catastrophic damage. We find that the effect of higher levels of strategic reasoning makes it harder to reach agreements in the model. This effect can be partially mitigated by restricting the range of initial bids in the bargaining process. In paper IV, we consider the arguments by Hanson and Bostrom about the Great Filter as an attempt to explain the Fermi Paradox. According to these arguments, finding extraterrestrial life on one planet should lower our expectations for humanity's prospects to progress far beyond our current technological capabilities. We model this claim as a Bayesian learning problem and examine the effect a single observation of life has in the model. We find that the conclusion of the argument depends critically on the choice of prior distribution. In paper V, we consider a model of agricultural markets and land-use competition between food and bioenergy crops. Agents in the model represent farmers who decide which crop to grow depending on predictors that give future price expectations. We model agents who can switch among predictors to make their decisions. We find that some predictor types can be concentrated on key parcels of land, which reduces volatility in crop prices for the system. We also examine several mechanisms that can bring price fluctuations in the system down and closer to a stable state.