Together with N Botta at the Potsdam Institute for Climate Impact Research and N Smallbone at the Chalmers FP Unit we are planning a research project around computer-aided policy-making and increasingly correct scientific computing. We have some preliminary results, but are looking for funding.
A crucial aspect of decision making is finding the right tradeoff between multiple conflicting objectives. For example, in a fusion reactor, disruption events can release undesirable levels of heat and electromagnetic forces. They can be controlled by injecting a combination of neon and deuterium into the reactor core. Certain combinations reduce the amount of heat more; others reduce the forces more. What is the best combination? The task is to explore the space of possible combinations and quantify their tradeoffs.
This is a multi-objective control problem. There is a system, a space of parameters known as control policies, and objectives to be met. There is no single best control policy, but some policies are Pareto-optimal: one can not improve one objective without worsening another. The goal is to identify robust Pareto-optimal controls.
Unfortunately, the control space is usually high-dimensional and the exploration cannot be based on data from empirical experiments: the relationship between controls and objectives has to be approximated via numerical simulations and the objectives associated with a control are affected by different kinds of uncertainties.
The purpose of this project is to build and apply a toolkit of open-source computational methods for exploring safe, fair and Pareto-optimal control policies for multi-objective control problems under uncertainty in close collaboration with domain experts from high energy physics and sustainability science.
Related work: OptiFun: Optimising fusion with generative programming, Bayesian optimization of fusion experiment simulations, Responsibility Under Uncertainty, The impact of uncertainty on optimal emission policies, Testing versus proving in climate impact research
A crucial aspect of decision making is finding the right tradeoff between multiple conflicting objectives. For example, in a fusion reactor, disruption events can release undesirable levels of heat and electromagnetic forces. They can be controlled by injecting a combination of neon and deuterium into the reactor core. Certain combinations reduce the amount of heat more; others reduce the forces more. What is the best combination? The task is to explore the space of possible combinations and quantify their tradeoffs.
This is a multi-objective control problem. There is a system, a space of parameters known as control policies, and objectives to be met. There is no single best control policy, but some policies are Pareto-optimal: one can not improve one objective without worsening another. The goal is to identify robust Pareto-optimal controls.
Unfortunately, the control space is usually high-dimensional and the exploration cannot be based on data from empirical experiments: the relationship between controls and objectives has to be approximated via numerical simulations and the objectives associated with a control are affected by different kinds of uncertainties.
The purpose of this project is to build and apply a toolkit of open-source computational methods for exploring safe, fair and Pareto-optimal control policies for multi-objective control problems under uncertainty in close collaboration with domain experts from high energy physics and sustainability science.
Related work: OptiFun: Optimising fusion with generative programming, Bayesian optimization of fusion experiment simulations, Responsibility Under Uncertainty, The impact of uncertainty on optimal emission policies, Testing versus proving in climate impact research