Free-form Regression Modelling

Service Description

FxC Reference Number : Climate-energy-technology simulation models are central to understanding and planning for transition to net zero CO2. They incorporate coupled simulation models of climate, energy supply and demand subdivided by energy technology, and market response. Providing expertise and consulting in such models to policy-makers is a niche sector within Ireland's consulting sector. However, such models are too large, unwieldy, and uninterpretable for direct use in consulting scenarios. The service offered here is effectively to create "a model of the model". The technology to be employed is an approach to free-form, interpretable regression modelling via metaheuristic methods ("genetic programming"). The end result of this modelling is a series of *equations* which are directly human-readable. Based on the results derived from the existing simulation models, these equations predict key quantities such as global cost at 2050 of shifts in global energy supply technologies, as a function of variables such as climate response, market sensitivity to climate change, and more. In contrast to linear regression models, the equations resulting from the service will be free-form, able to capture the non-linear relationships in the data. In contrast to modern non-linear regression models such as neural networks and gradient-boosted trees, the equations resulting from the service will be highly readable and human-interpretable. Therefore, the results of the service will be suitable for direct use in consulting services. The service includes: (i) One-to-one sessions to develop goals and understand data; (ii) Many large computational runs to establish baselines of performance with standard regression methods; (iii) Many large computational runs to develop the *genetic programming* regression models, adhering to best practice for measuring quality and ensuring model robustness and interpretability; (iv) Presentation and Q&A to convey the results, including quality, robustness, interpretability, key influential variables, variance of multiple runs, and risks / limitations; (v) Discussion on possible next steps such as data visualisation or further regression modelling; (vi) Creation of a report with publication-quality plots and tables, and open-source code and results data. An agile development approach will be used. A baseline result will be presented in the first week in order to make the service concrete and enable discussion. Each further i