Risk CDT - Uncertainty and Extreme Outcomes in Agent-based Models of Complex Systems

PLEASE APPLY ONLINE TO THE SCHOOL OF ENGINEERING, PROVIDING THE PROJECT TITLE, NAME OF THE PRIMARY SUPERVISOR AND SELECT THE PROGRAMME CODE "EGPR" (PHD - SCHOOL OF ENGINEERING)

This is a project within the multi-disciplinary EPSRC and ESRC Centre for Doctoral Training (CDT) on Quantification and Management of Risk & Uncertainty in Complex Systems & Environments, within the Institute for Risk and Uncertainty. The studentship is granted for 4 years and includes, in the first year, a Master in Decision Making under Risk & Uncertainty. The project includes extensive collaboration with prime industry to build an optimal basis for employability.

Probabilistic risk analysis based on sophisticated Monte Carlo methods is a well-developed field, yet predictive failures are more common in these probabilistic simulations than they should be. For instance, NASA estimated the risk of catastrophic failure of the Shuttle to be 1/10,000 per flight, but its observed failure rate was about 1 per 70 flights. Flood control authorities underestimate risks of major floods and miscommunicate uncertainties about flooding risks. Observed failures and near-misses found at some nuclear power plants reveal understatement in their assessed risks. Failure cascades in electricity distribution systems are more common than they are forecasted to be. Underestimating risks in the financial industry precipitated the recent global recession. Biological consequences of the introduction of exotic organisms and diseases into new environments are often completely unanticipated. These are not random rare events in a discipline of risk analysis that estimates and controls risks well overall. They seem instead to be the result of pervasive and systemic errors that undermine the credibility of many modelling and simulation efforts in risk analysis.

While one cause of these predictive failures is using analytic methods under assumptions that are not justified by empirical evidence, another important cause may be the use of Monte Carlo simulation itself as the analytical tool. Sampling-based strategies often cannot meet fundamental analytical needs, especially in exploring tail risks of extreme behaviours and identifying worst-case outcomes, both of which are critical in engineering planning. Monte Carlo techniques are well suited to identifying average and typical behaviours of complex systems, but failures are not usually typical behaviours. These limitations are exacerbated when the dimensionality of the problem is very high, and in situations where volitional decisions by humans are involved. Standard Monte Carlo simulations are not very good at identifying weak points in such complex systems, even though malevolent actors such as terrorists, arsonists, and economic competitors often readily find those weak points. If the simulation predictions miss things that can happen, or seriously underestimate their chances of happening, the analyses are really only Potemkin Villages of proper assessments. They fail to show us just how bad things could be, and thus we are unprepared when exigencies arise. Even worse, they create a pretence of serious analysis where there is none, which precludes expenditures of effort that could otherwise actually be useful.

Analytical tools are needed that overcome the deficiencies of sampling-based strategies. In this project, we will develop such tools for use in analyses of probabilistic agent-based models using methods that computer scientists call “automatically verified computations”. These methods are guaranteed to enclose the actual numerical results and rigorously contain the effects of representational error, computational round-off error, measurement imprecision about input data, and modelling uncertainty arising from doubt about distribution shape and inter-variable dependency. The proposed tools will make use of modern approaches including robust Bayes analysis developed over the last few years that focus on a fundamental distinction between epistemic uncertainty and aleatory uncertainty and how this distinction manifests itself in inferences and calculations. The developed tools will be tested in case studies by application to models of disease epidemics in biological populations. The case studies will include epidemics unfolding naturally, epidemics met with epidemiological control measures, and epidemics sparked in terrorism scenarios. Agent-based modelling is employed for a host of engineering problems, especially for systems in which many independent or semi-independent entities interact, such as biological populations, sales forces, consumer markets, electorates, armies, social communities, and gaming networks, among others.

This project will be computationally intensive. Supercomputer time may be useful. As appropriate, methods developed will be incorporated into or made otherwise accessible to users of COSSAN-X.