While machine learning has recently had dramatic successes, there is a large class of problems that it will never be able to address on its own. To test a policy proposal, for example, often requires understanding a counterfactual scenario that has never existed in the past, and that may fundamentally alter the statistical properties of data in the future. Simulation models provide an alternative, in which one incorporates much stronger prior knowledge about structure and causal interaction. Making such models quantitatively accurate enough that they can be trusted for policy analysis poses difficult challenges for parameter estimation and initialization. Simulation models are usually formulated in terms of micro-states, such as individual households, whereas measurements are often only available at an aggregate level, such as GDP or unemployment. To run a simulation as a time series model to forecast macro-states requires initializing the micro-states of the model to match macroscopic data, and can create severe problems if not done carefully.
Borrowing from methods of data assimilation used in meteorology, we introduce a new method for solving this problem. I will review the literature on parameter estimation for simulation models and discuss the relevant challenges and the opportunities to create a fusion with machine learning. My main point is that the problems of parameter estimation and initialization for simulation models are at the cutting edge of data science. These problems are ripe to be solved, and their solution will catapult simulation science to a level of usefulness similar to that of machine learning. Work in this area should figure prominently on the agenda of The Alan Turing Institute.
Doyne Farmer is Director of the Complexity Economics program at the Institute for New Economic Thinking at the Oxford Martin School, Professor in the Mathematical Institute at the University of Oxford, and an External Professor at the Santa Fe Institute. His current research is in economics and finance, including agent-based modeling, financial instability and technological progress. He was a founder of Prediction Company, a quantitative automated trading firm that was sold to the United Bank of Switzerland in 2006. His past research includes complex systems, dynamical systems theory, time series analysis and theoretical biology. During the eighties he was an Oppenheimer Fellow and the founder of the Complex Systems Group at Los Alamos National Laboratory. While a graduate student in the 70’s he built the first wearable digital computer, which was successfully used to predict the game of roulette.
Add comment