TITLE:  Learning in integrated optimization models of climate change and economy

STUDENT:  Soheil Shayegh

ADVISOR:  Dr. Valerie Thomas

SUMMARY:  

In this research we study learning in optimization models with uncertainty. The focus is on applications of adaptive stochastic programming in integrated assessment modeling of climate change and economy.

In the first part of the thesis we investigate finite horizon stochastic models with continuous state and action space. We develop an approximation technique for calculating the value function endogenously and in an online fashion. We introduce a multistep lookahead value iteration algorithm to update the parameters of the approximation function. We apply this technique to a dynamic integrated assessment model of climate change and economy. We present the results of our analysis in both deterministic and stochastic cases. We show that the multistep ahead approximation technique is capable of approximating the value function fast and converging to its true value. We study the optimal behavior of the model under uncertainty from climate parameters and shocks. We show how a stochastic model can inform climate mitigation decisions when there is risk of an extreme event, with larger actions before an extreme event, and potentially smaller actions after.

The second part of the thesis looks at stochastic models with correlated uncertainties. We address the correlation between uncertain temperature increase and contingent climate extreme events. As temperature increases so does the chance of having an extreme event. The occurrence of an extreme event provides information about the distribution of the temperature controlling parameter. We address this correlation through Bayesian inference. We apply an automatic update to our previous algorithm and show how the posterior distribution of the control parameter is updated as we receive more information about the realization of the dependent parameter through simulation. We show how the optimal solution varies as learning happens over time and updates the prior probability distributions.

In the third part of the thesis, we explore a different aspect of learning: learning-by-doing. We consider the global electricity generation expansion planning problem with or without climate regulations. We model a spectrum of electricity generation technologies with different emissions, costs, and learning rates. We show how the optimal mix of technologies varies under different climate regulatory regimes. The optimal solution also depends on the learning assumption and the social time preference factor (discounting). We see that in both cases, with or without climate regulations, new technologies enter the optimal portfolio if the learning is applied to the model. We also compare the results in terms of impact on the global mean surface temperature and show that early investment in new low carbon technologies can lower the long-term temperature trajectory.